feat(webui): add computer use controls

Adds two opt-in agent tools for controlling a computer:
- computer_use: pixel-based (screenshot + mouse/keyboard) via a desktop
  (pyautogui) or browser (playwright) backend.
- browser: DOM/accessibility-based web automation (act by element ref),
  reliable across ANY tool-calling model, not just vision/CU-trained ones.

Core enabler in the runner: a tool may return image content blocks, which
are split out and delivered to the model as a follow-up user message
(_split_tool_result_media), so screenshots reach any vision provider
(e.g. via OpenRouter/openai-compat) without provider-specific code.

Both tools are OFF by default (tools.computerUse.enable / tools.browser.enable),
are not exposed to subagents, and the browser tool supports an allowed_domains
allowlist. Heavy deps (pyautogui/pillow/playwright) are an optional [computer-use] extra.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/architecture.md

@@ -125,6 +125,7 @@ Important files:
 | Shell execution | `nanobot/agent/tools/shell.py` |
 | Filesystem tools | `nanobot/agent/tools/filesystem.py` |
 | Web search/fetch | `nanobot/agent/tools/web.py` |
+| Browser and computer use | `nanobot/agent/tools/browser_tool.py`, `nanobot/agent/tools/computer_use.py` |
 | MCP tools | `nanobot/agent/tools/mcp.py` |
 | Cron | `nanobot/agent/tools/cron.py`, `nanobot/cron/` |
 | Image generation | `nanobot/agent/tools/image_generation.py` |
@@ -188,7 +189,7 @@ Security-sensitive code paths include:
 |---|---|
 | Workspace scope | `nanobot/security/workspace_access.py`, `nanobot/security/workspace_policy.py` |
 | Shell sandboxing | `nanobot/agent/tools/shell.py` |
-| SSRF/network checks | `nanobot/security/network.py`, `nanobot/agent/tools/web.py` |
+| SSRF/network checks | `nanobot/security/network.py`, `nanobot/agent/tools/web.py`, `nanobot/agent/tools/computer_use_backends/browser_playwright.py` |
 | PTH guard and CLI startup security | `nanobot/security/` and CLI entrypoints |
 | Channel access control | channel config in `nanobot/channels/*.py` |
 

docs/configuration.md

@@ -42,6 +42,7 @@ the focused guides first and come back here for exact fields and defaults.
 | Add fallback chains | [Model Fallbacks](#model-fallbacks) |
 | Configure voice transcription | [Transcription Settings](#transcription-settings) |
 | Tune channel defaults | [Channel Settings](#channel-settings) |
+| Enable browser or desktop control | [Browser and Computer Use](#browser-and-computer-use) |
 | Configure web search and fetch | [Web Tools](#web-tools) |
 | Enable image generation | [Image Generation](#image-generation) |
 | Add MCP servers | [MCP](#mcp-model-context-protocol) |
@@ -1670,6 +1671,62 @@ When a channel `send()` raises, nanobot retries at the channel-manager layer. By
 >
 > If a channel is completely unreachable, nanobot cannot notify the user through that same channel. Watch logs for `Failed to send to {channel} after N attempts` to spot persistent delivery failures.
 
+## Browser and Computer Use
+
+Browser and desktop control are optional and disabled by default. Install their runtime first:
+
+```bash
+pip install 'nanobot-ai[computer-use]'
+playwright install chromium
+```
+
+For normal web interaction, prefer the DOM-based `browser` tool. It gives the model numbered
+element references and works without vision. Use `computer_use` when the model must see and act
+on pixels; its `desktop` backend controls the real local machine, while its `browser` backend
+controls an isolated Playwright page.
+
+```json
+{
+  "tools": {
+    "browser": {
+      "enable": true,
+      "allowedDomains": ["example.com"]
+    },
+    "computerUse": {
+      "enable": false,
+      "backend": "desktop"
+    }
+  }
+}
+```
+
+| Option | Default | Description |
+|---|---|---|
+| `tools.browser.enable` | `false` | Register the DOM-based `browser` tool |
+| `tools.browser.allowedDomains` | `[]` | Optional top-level navigation allowlist; entries include subdomains |
+| `tools.browser.includeScreenshot` | `false` | Attach a screenshot after browser actions |
+| `tools.browser.maxSessions` | `8` | Maximum retained browser sessions; least-recently-used state is closed first |
+| `tools.computerUse.enable` | `false` | Register pixel-based `computer_use` |
+| `tools.computerUse.backend` | `"desktop"` | `"desktop"` or `"browser"` |
+| `tools.computerUse.allowedDomains` | `[]` | Navigation allowlist for the browser backend |
+| `tools.computerUse.targetWidth` / `targetHeight` | `1280` / `800` | Maximum screenshot dimensions exposed to the model |
+| `tools.computerUse.maxSessions` | `8` | Maximum retained sessions for the browser backend |
+
+Each nanobot session gets separate browser state. Browser HTTP and WebSocket traffic passes
+through the shared SSRF policy; local, private, link-local, and metadata targets are blocked
+unless explicitly permitted with `tools.ssrfWhitelist`. When `maxSessions` is reached, the
+least-recently-used browser state is closed. `file:` URLs are not accepted.
+
+> [!IMPORTANT]
+> Browser URL checks are defense in depth, not an egress sandbox: Chromium performs its own DNS
+> resolution after validation. Use OS/container network isolation when browsing hostile pages.
+
+> [!WARNING]
+> The desktop backend can click, type, and change state outside the workspace. Enabling it is an
+> explicit trust decision: use a trusted model and input source, and run nanobot in a disposable
+> OS account or VM when unattended. The workspace restriction is not an OS sandbox. Desktop text
+> input supports ASCII key events; use the browser backend when Unicode text input is required.
+
 ## Web Tools
 
 nanobot incorporates basic tools for accessing the web. These include searching via APIs, and fetching arbitrary web pages in Markdown format. They are enabled by default, and can be configured in `~/.nanobot/config.json` under `tools.web`.

nanobot/agent/context_governance.py

@@ -33,6 +33,11 @@ COMPACTABLE_TOOLS = frozenset({
     "read_file", "exec", "grep", "find_files",
     "web_search", "web_fetch", "list_dir", "list_exec_sessions",
 })
+VISUAL_TOOLS = frozenset({"browser", "computer_use"})
+STALE_SCREENSHOT_PLACEHOLDER = {
+    "type": "text",
+    "text": "[Earlier screenshot omitted; use the latest screenshot from this tool.]",
+}
 # read_file is the recovery path for persisted results; exempting it prevents persist->read->persist loops.
 TOOL_RESULT_OFFLOAD_EXEMPT_TOOLS = frozenset({"read_file"})
 BACKFILL_CONTENT = "[Tool result unavailable — call was interrupted or lost]"
@@ -41,6 +46,12 @@ PLACEHOLDER_TEXTS = frozenset({
 })
 
 
+def _is_image_block(value: object) -> bool:
+    if not isinstance(value, dict):
+        return False
+    return cast(dict[str, Any], value).get("type") in {"image_url", "input_image"}
+
+
 def _tool_call_name_is_valid(tool_call: Any) -> bool:
     """Whether a persisted OpenAI-style tool_call carries a usable name.
 
@@ -84,6 +95,10 @@ class ContextGovernor:
         updated = self.drop_orphan_tool_results(updated)
         updated = self.backfill_missing_tool_results(updated)
         updated = self.apply_tool_result_budget(config, updated)
+        updated = self.drop_stale_visual_tool_images(
+            updated,
+            start_index=config.inflight_start_index,
+        )
         updated = self.compact_inflight_overflow(config, updated, compacted_tool_call_ids)
         updated = self.snip_history(config, updated)
         updated = self.drop_orphan_tool_results(updated)
@@ -326,6 +341,35 @@ class ContextGovernor:
                 updated[idx]["content"] = normalized
         return updated
 
+    @staticmethod
+    def drop_stale_visual_tool_images(
+        messages: list[dict[str, Any]],
+        *,
+        start_index: int,
+    ) -> list[dict[str, Any]]:
+        """Keep only the latest in-flight screenshot from each visual tool."""
+        seen: set[str] = set()
+        updated = messages
+        for idx in range(len(messages) - 1, start_index - 1, -1):
+            message = messages[idx]
+            name = str(message.get("name") or "")
+            content = message.get("content")
+            if message.get("role") != "tool" or name not in VISUAL_TOOLS:
+                continue
+            if not isinstance(content, list):
+                continue
+            content_blocks = cast(list[object], content)
+            blocks = [block for block in content_blocks if not _is_image_block(block)]
+            if len(blocks) == len(content_blocks):
+                continue
+            if name not in seen:
+                seen.add(name)
+                continue
+            if updated is messages:
+                updated = [dict(item) for item in messages]
+            updated[idx]["content"] = [dict(STALE_SCREENSHOT_PLACEHOLDER), *blocks]
+        return updated
+
     def compact_inflight_overflow(
         self,
         config: ContextGovernanceConfig,

nanobot/agent/loop.py

@@ -1397,6 +1397,7 @@ class AgentLoop:
         cleanup_steps = (
             self.subagents.close,
             self._exec_session_manager.close_all,
+            *(() if not hasattr(self, "tools") else (self.tools.close,)),
             lambda: agent_context.close_mcp(self),
         )
         for cleanup in cleanup_steps:

nanobot/agent/tools/base.py

@@ -220,6 +220,10 @@ class Tool(ABC):
         """Return optional per-turn prompt context owned by this tool."""
         return None
 
+    async def close(self) -> None:
+        """Release resources owned by the tool. Safe to call repeatedly."""
+        return None
+
     @abstractmethod
     async def execute(self, **kwargs: Any) -> Any:
         """Run the tool; return content, or ``ToolResult.error(...)`` for failures."""

nanobot/agent/tools/browser_tool.py

@@ -0,0 +1,280 @@
+"""DOM-based browser automation by element reference."""
+
+# pyright: reportIncompatibleMethodOverride=false
+
+from __future__ import annotations
+
+import asyncio
+from typing import Any
+
+from pydantic import Field
+
+from nanobot.agent.tools.base import Tool, tool_parameters
+from nanobot.agent.tools.computer_use_backends.base import SessionBackendPool
+from nanobot.agent.tools.context import ToolContext
+from nanobot.agent.tools.schema import (
+    BooleanSchema,
+    IntegerSchema,
+    StringSchema,
+    tool_parameters_schema,
+)
+from nanobot.config_base import Base
+from nanobot.utils.helpers import build_image_content_blocks
+
+_ACTIONS = [
+    "navigate",
+    "snapshot",
+    "click",
+    "type",
+    "select",
+    "scroll",
+    "key",
+    "back",
+    "read_text",
+]
+
+
+class BrowserToolConfig(Base):
+    """browser (DOM) tool configuration."""
+
+    enable: bool = False
+    start_url: str = "about:blank"
+    headless: bool = True
+    width: int = Field(default=1280, ge=320, le=4096)
+    height: int = Field(default=800, ge=240, le=4096)
+    allowed_domains: list[str] = Field(default_factory=list)
+    include_screenshot: bool = False
+    max_elements: int = Field(default=200, ge=1, le=1000)
+    max_sessions: int = Field(default=8, ge=1, le=64)
+
+
+def _format_elements(elements: list[dict[str, Any]]) -> str:
+    if not elements:
+        return "Interactive elements: (none found — try scrolling or read_text)"
+    lines: list[str] = []
+    for e in elements:
+        tag = str(e.get("tag") or "")
+        typ = str(e.get("type") or "")
+        label = tag + (f"[{typ}]" if typ else "")
+        line = f"[{e.get('ref')}] {label}"
+        name = str(e.get("name") or "").strip()
+        if name:
+            line += f' "{name}"'
+        href = str(e.get("href") or "")
+        if href and tag == "a":
+            line += f" -> {href[:60]}"
+        lines.append(line)
+    return "Interactive elements (act with the [ref] number):\n" + "\n".join(lines)
+
+
+@tool_parameters(
+    tool_parameters_schema(
+        action=StringSchema("The action to perform.", enum=_ACTIONS),
+        ref=IntegerSchema(
+            description="Element ref number from the latest snapshot (click/type/select).",
+            minimum=1,
+            nullable=True,
+        ),
+        text=StringSchema(
+            "Text to type (action=type) or key/combo like 'Enter'/'ctrl+a' (action=key).",
+            nullable=True,
+        ),
+        url=StringSchema("URL to open (action=navigate).", nullable=True),
+        value=StringSchema("Option value/label to choose (action=select).", nullable=True),
+        submit=BooleanSchema(description="Press Enter after typing (action=type).", nullable=True),
+        scroll_direction=StringSchema(
+            "Scroll direction (action=scroll).", enum=["up", "down", "left", "right"], nullable=True
+        ),
+        scroll_amount=IntegerSchema(
+            description="Scroll clicks (action=scroll).",
+            minimum=1,
+            maximum=100,
+            nullable=True,
+        ),
+        required=["action"],
+    )
+)
+class BrowserTool(Tool):
+    """Browse and act on web pages by element ref (DOM-based, works with any model)."""
+
+    _scopes = {"core"}
+
+    name = "browser"  # pyright: ignore[reportIncompatibleMethodOverride, reportAssignmentType]
+    description = (  # pyright: ignore[reportIncompatibleMethodOverride, reportAssignmentType]
+        "Control a web browser by acting on page elements by their [ref] number. "
+        "Each call returns the current page URL plus a fresh numbered list of the page's "
+        "interactive elements; pick a [ref] to click/type/select — no pixel coordinates "
+        "needed. A page may already be open: call 'snapshot' FIRST to see it. Only use "
+        "'navigate' for a specific URL you were explicitly given — never guess a URL. "
+        "Move between pages by clicking links/buttons via their [ref]. Use 'read_text' to "
+        "read page text. Re-read the element list after each action; refs are reassigned."
+    )
+
+    config_key = "browser"
+
+    @classmethod
+    def config_cls(cls) -> type[BrowserToolConfig]:
+        return BrowserToolConfig
+
+    @classmethod
+    def enabled(cls, ctx: ToolContext) -> bool:
+        return bool(ctx.config.browser.enable)
+
+    @classmethod
+    def create(cls, ctx: ToolContext) -> Tool:
+        return cls(ctx.config.browser)
+
+    def __init__(
+        self,
+        config: BrowserToolConfig | None = None,
+        *,
+        backend_impl: Any = None,
+    ) -> None:
+        self.config = config or BrowserToolConfig()
+        runtime = None
+        if backend_impl is None:
+            from nanobot.agent.tools.computer_use_backends.browser_playwright import BrowserRuntime
+            runtime = BrowserRuntime(headless=self.config.headless)
+        self._runtime = runtime
+        self._execution_lock = asyncio.Lock()
+        self._backends = SessionBackendPool(
+            self._make_backend,
+            backend_impl,
+            max_backends=self.config.max_sessions,
+            finalizer=runtime.close if runtime is not None else None,
+        )
+
+    @property
+    def read_only(self) -> bool:
+        return False
+
+    @property
+    def exclusive(self) -> bool:
+        return True
+
+    def _make_backend(self) -> Any:
+        from nanobot.agent.tools.computer_use_backends.browser_playwright import BrowserBackend
+        return BrowserBackend(
+            width=self.config.width,
+            height=self.config.height,
+            start_url=self.config.start_url,
+            allowed_domains=self.config.allowed_domains,
+            runtime=self._runtime,
+        )
+
+    @staticmethod
+    def _req_ref(params: dict[str, Any], action: str) -> Any:
+        ref = params.get("ref")
+        if ref is None:
+            raise ValueError(f"action '{action}' requires an element 'ref' from the snapshot")
+        return ref
+
+    async def _dispatch(self, backend: Any, action: str, p: dict[str, Any]) -> tuple[str, str | None]:
+        """Return (status, direct_text). If direct_text is set, it is returned as-is
+        (no snapshot appended)."""
+        if action == "navigate":
+            url = p.get("url")
+            if not url:
+                raise ValueError("action 'navigate' requires 'url'")
+            await backend.navigate(str(url))
+            return f"Navigated to {url}", None
+
+        if action == "snapshot":
+            return "Snapshot of the current page", None
+
+        if action == "click":
+            ref = self._req_ref(p, action)
+            await backend.click_ref(ref)
+            return f"Clicked element [{ref}]", None
+
+        if action == "type":
+            ref = self._req_ref(p, action)
+            text = p.get("text")
+            if text is None:
+                raise ValueError("action 'type' requires 'text'")
+            submit = bool(p.get("submit"))
+            await backend.fill_ref(ref, str(text), submit=submit)
+            return f"Typed into [{ref}]" + (" and pressed Enter" if submit else ""), None
+
+        if action == "select":
+            ref = self._req_ref(p, action)
+            value = p.get("value")
+            if value is None:
+                raise ValueError("action 'select' requires 'value'")
+            await backend.select_ref(ref, str(value))
+            return f"Selected '{value}' in [{ref}]", None
+
+        if action == "scroll":
+            direction = str(p.get("scroll_direction") or "down").lower()
+            if direction not in ("up", "down", "left", "right"):
+                raise ValueError("'scroll_direction' must be up/down/left/right")
+            await backend.scroll_page(direction, int(p.get("scroll_amount") or 3))
+            return f"Scrolled {direction}", None
+
+        if action == "key":
+            combo = p.get("text")
+            if not combo:
+                raise ValueError("action 'key' requires 'text' (e.g. 'Enter')")
+            await backend.key(str(combo))
+            return f"Pressed {combo}", None
+
+        if action == "back":
+            await backend.go_back()
+            return "Navigated back", None
+
+        if action == "read_text":
+            txt = await backend.read_text()
+            return "", f"Page text:\n{txt}"
+
+        raise ValueError(f"unknown action '{action}'")
+
+    async def execute(self, action: str | None = None, **kwargs: Any) -> Any:
+        async with self._execution_lock:
+            return await self._execute(action, **kwargs)
+
+    async def _execute(self, action: str | None = None, **kwargs: Any) -> Any:
+        action = (action or "").strip()
+        if action not in _ACTIONS:
+            return f"Error: unknown action '{action}'. Valid actions: {', '.join(_ACTIONS)}"
+
+        try:
+            backend = await self._backends.get()
+        except ImportError as exc:
+            return f"Error: {exc}"
+        except Exception as exc:
+            return f"Error: could not initialize browser backend: {type(exc).__name__}: {exc}"
+
+        try:
+            status, direct = await self._dispatch(backend, action, kwargs)
+            if blocked := getattr(backend, "pop_blocked_navigation", lambda: None)():
+                raise ValueError(f"navigation was blocked: {blocked}")
+        except ValueError as exc:
+            return f"Error: {exc}"
+        except Exception as exc:
+            return f"Error executing browser '{action}': {type(exc).__name__}: {exc}"
+
+        if direct is not None:
+            return direct
+
+        try:
+            elements = await backend.dom_snapshot(self.config.max_elements)
+            snapshot = _format_elements(elements)
+        except Exception as exc:
+            snapshot = f"(could not read page elements: {type(exc).__name__}: {exc})"
+        try:
+            current = await backend.current_url()
+        except Exception:
+            current = ""
+        header = f"{status}\nCurrent page: {current}" if current else status
+        text_out = f"{header}\n\n{snapshot}"
+
+        if self.config.include_screenshot:
+            try:
+                png = await backend.screenshot()
+                return build_image_content_blocks(png, "image/png", "", text_out)
+            except Exception:
+                return text_out
+        return text_out
+
+    async def close(self) -> None:
+        await self._backends.close()

nanobot/agent/tools/computer_use.py

@@ -0,0 +1,327 @@
+"""Screenshot-based computer control."""
+
+# pyright: reportIncompatibleMethodOverride=false
+
+from __future__ import annotations
+
+import asyncio
+import io
+from typing import Any, Literal
+
+from pydantic import Field
+
+from nanobot.agent.tools.base import Tool, tool_parameters
+from nanobot.agent.tools.computer_use_backends.base import SessionBackendPool
+from nanobot.agent.tools.context import ToolContext
+from nanobot.agent.tools.schema import (
+    IntegerSchema,
+    NumberSchema,
+    StringSchema,
+    tool_parameters_schema,
+)
+from nanobot.config_base import Base
+from nanobot.utils.helpers import build_image_content_blocks
+
+_ACTIONS = [
+    "screenshot",
+    "left_click",
+    "right_click",
+    "middle_click",
+    "double_click",
+    "triple_click",
+    "mouse_move",
+    "left_click_drag",
+    "scroll",
+    "type",
+    "key",
+    "wait",
+    "navigate",
+]
+
+_CLICK_BUTTONS = {
+    "left_click": "left",
+    "double_click": "left",
+    "triple_click": "left",
+    "right_click": "right",
+    "middle_click": "middle",
+}
+_CLICK_COUNTS = {"double_click": 2, "triple_click": 3}
+
+_MAX_WAIT_S = 10.0
+
+
+class ComputerUseToolConfig(Base):
+    """computer_use tool configuration."""
+
+    enable: bool = False
+    backend: Literal["desktop", "browser"] = "desktop"
+    target_width: int = Field(default=1280, ge=320, le=4096)
+    target_height: int = Field(default=800, ge=240, le=4096)
+    allowed_domains: list[str] = Field(default_factory=list)
+    start_url: str = "about:blank"
+    headless: bool = True
+    max_sessions: int = Field(default=8, ge=1, le=64)
+
+
+def _fit_size(width: int, height: int, max_width: int, max_height: int) -> tuple[int, int]:
+    if width <= 0 or height <= 0:
+        return max(1, max_width), max(1, max_height)
+    scale = min(max_width / width, max_height / height, 1.0)
+    return max(1, round(width * scale)), max(1, round(height * scale))
+
+
+def _scale_point(
+    x: int,
+    y: int,
+    source: tuple[int, int],
+    target: tuple[int, int],
+) -> tuple[int, int]:
+    width, height = source
+    target_width, target_height = target
+    real_x = round(x * width / target_width) if target_width else x
+    real_y = round(y * height / target_height) if target_height else y
+    return (
+        max(0, min(real_x, max(0, width - 1))),
+        max(0, min(real_y, max(0, height - 1))),
+    )
+
+
+@tool_parameters(
+    tool_parameters_schema(
+        action=StringSchema("The action to perform.", enum=_ACTIONS),
+        x=IntegerSchema(
+            description="X coordinate in the pixel space of the screenshot you were last shown.",
+            nullable=True,
+        ),
+        y=IntegerSchema(
+            description="Y coordinate in the pixel space of the screenshot you were last shown.",
+            nullable=True,
+        ),
+        text=StringSchema(
+            "Text to type (action=type; desktop supports ASCII), or a key/combo like "
+            "'ctrl+s' or 'Enter' (action=key).",
+            nullable=True,
+        ),
+        scroll_direction=StringSchema(
+            "Scroll direction (action=scroll).", enum=["up", "down", "left", "right"], nullable=True
+        ),
+        scroll_amount=IntegerSchema(
+            description="Number of scroll clicks (action=scroll).",
+            minimum=1,
+            maximum=100,
+            nullable=True,
+        ),
+        duration=NumberSchema(
+            description="Seconds to wait (action=wait).",
+            minimum=0,
+            maximum=_MAX_WAIT_S,
+            nullable=True,
+        ),
+        url=StringSchema("URL to open (action=navigate, browser backend only).", nullable=True),
+        required=["action"],
+    )
+)
+class ComputerUseTool(Tool):
+    """Control a computer (desktop or browser) by looking at screenshots and acting."""
+
+    _scopes = {"core"}  # never exposed to subagents — security-sensitive
+
+    name = "computer_use"  # pyright: ignore[reportIncompatibleMethodOverride, reportAssignmentType]
+    description = (  # pyright: ignore[reportIncompatibleMethodOverride, reportAssignmentType]
+        "Control a computer via screenshots and mouse/keyboard. Each call performs ONE "
+        "action and returns a fresh screenshot of the resulting screen. Coordinates (x, y) "
+        "are in the pixel space of the screenshot you were last shown (top-left is 0,0). "
+        "The 'browser' backend additionally supports the 'navigate' action. Always start "
+        "with a 'screenshot' to see the screen, then act based on what you observe; after "
+        "each action re-check the new screenshot before the next step."
+    )
+
+    config_key = "computer_use"
+
+    @classmethod
+    def config_cls(cls) -> type[ComputerUseToolConfig]:
+        return ComputerUseToolConfig
+
+    @classmethod
+    def enabled(cls, ctx: ToolContext) -> bool:
+        return bool(ctx.config.computer_use.enable)
+
+    @classmethod
+    def create(cls, ctx: ToolContext) -> Tool:
+        return cls(ctx.config.computer_use)
+
+    def __init__(
+        self,
+        config: ComputerUseToolConfig | None = None,
+        *,
+        backend_impl: Any = None,
+    ) -> None:
+        self.config = config or ComputerUseToolConfig()
+        runtime = None
+        if backend_impl is None and self.config.backend == "browser":
+            from nanobot.agent.tools.computer_use_backends.browser_playwright import BrowserRuntime
+            runtime = BrowserRuntime(headless=self.config.headless)
+        self._runtime = runtime
+        self._execution_lock = asyncio.Lock()
+        self._backends = SessionBackendPool(
+            self._make_backend,
+            backend_impl,
+            max_backends=1 if self.config.backend == "desktop" else self.config.max_sessions,
+            finalizer=runtime.close if runtime is not None else None,
+        )
+
+    @property
+    def read_only(self) -> bool:
+        return False
+
+    @property
+    def exclusive(self) -> bool:
+        # Stateful single environment; must not run alongside other tools.
+        return True
+
+    def _make_backend(self) -> Any:
+        if self.config.backend == "browser":
+            from nanobot.agent.tools.computer_use_backends.browser_playwright import BrowserBackend
+            return BrowserBackend(
+                width=self.config.target_width,
+                height=self.config.target_height,
+                start_url=self.config.start_url,
+                allowed_domains=self.config.allowed_domains,
+                runtime=self._runtime,
+            )
+        from nanobot.agent.tools.computer_use_backends.desktop_pyautogui import DesktopBackend
+        return DesktopBackend()
+
+    @staticmethod
+    def _downscale_png(png: bytes, target: tuple[int, int]) -> bytes:
+        try:
+            from PIL import Image  # noqa: PLC0415
+        except Exception as exc:
+            raise ImportError(
+                "Pillow is required for computer_use. Install: pip install 'nanobot-ai[computer-use]'"
+            ) from exc
+        tw, th = target
+        with Image.open(io.BytesIO(png)) as img:
+            if (img.width, img.height) == (tw, th):
+                return png
+            resized = img.convert("RGB").resize((tw, th))  # pyright: ignore[reportUnknownMemberType]
+            out = io.BytesIO()
+            resized.save(out, format="PNG")
+            return out.getvalue()
+
+    async def _dispatch(
+        self,
+        backend: Any,
+        action: str,
+        params: dict[str, Any],
+        source: tuple[int, int],
+        target: tuple[int, int],
+    ) -> str:
+        def _xy() -> tuple[int, int]:
+            x, y = params.get("x"), params.get("y")
+            if x is None or y is None:
+                raise ValueError(f"action '{action}' requires integer 'x' and 'y'")
+            return _scale_point(int(x), int(y), source, target)
+
+        if action == "screenshot":
+            return "Took a screenshot"
+
+        if action == "wait":
+            duration = params.get("duration")
+            secs = 1.0 if duration is None else float(duration)
+            secs = max(0.0, min(secs, _MAX_WAIT_S))
+            await asyncio.sleep(secs)
+            return f"Waited {secs:g}s"
+
+        if action in _CLICK_BUTTONS:
+            rx, ry = _xy()
+            await backend.click(rx, ry, _CLICK_BUTTONS[action], _CLICK_COUNTS.get(action, 1))
+            return f"{action} at ({rx}, {ry})"
+
+        if action == "mouse_move":
+            rx, ry = _xy()
+            await backend.move(rx, ry)
+            return f"Moved to ({rx}, {ry})"
+
+        if action == "left_click_drag":
+            rx, ry = _xy()
+            await backend.drag(rx, ry)
+            return f"Dragged to ({rx}, {ry})"
+
+        if action == "scroll":
+            rx, ry = _xy()
+            direction = str(params.get("scroll_direction") or "down").lower()
+            if direction not in ("up", "down", "left", "right"):
+                raise ValueError("'scroll_direction' must be up/down/left/right")
+            amount = int(params.get("scroll_amount") or 3)
+            await backend.scroll(rx, ry, direction, amount)
+            return f"Scrolled {direction} by {amount} at ({rx}, {ry})"
+
+        if action == "type":
+            text = params.get("text")
+            if not text:
+                raise ValueError("action 'type' requires 'text'")
+            await backend.type_text(str(text))
+            return f"Typed {len(str(text))} characters"
+
+        if action == "key":
+            combo = params.get("text")
+            if not combo:
+                raise ValueError("action 'key' requires 'text' (e.g. 'ctrl+s')")
+            await backend.key(str(combo))
+            return f"Pressed {combo}"
+
+        if action == "navigate":
+            url = params.get("url")
+            if not url:
+                raise ValueError("action 'navigate' requires 'url'")
+            await backend.navigate(str(url))
+            return f"Navigated to {url}"
+
+        raise ValueError(f"unknown action '{action}'")
+
+    async def execute(self, action: str | None = None, **kwargs: Any) -> Any:
+        async with self._execution_lock:
+            return await self._execute(action, **kwargs)
+
+    async def _execute(self, action: str | None = None, **kwargs: Any) -> Any:
+        action = (action or "").strip()
+        if action not in _ACTIONS:
+            return f"Error: unknown action '{action}'. Valid actions: {', '.join(_ACTIONS)}"
+
+        try:
+            backend = await self._backends.get()
+            real_w, real_h = await backend.dimensions()
+        except ImportError as exc:
+            return f"Error: {exc}"
+        except Exception as exc:
+            return f"Error: could not initialize computer_use backend: {type(exc).__name__}: {exc}"
+
+        source = (real_w, real_h)
+        target = _fit_size(real_w, real_h, self.config.target_width, self.config.target_height)
+
+        try:
+            status = await self._dispatch(backend, action, kwargs, source, target)
+            if blocked := getattr(backend, "pop_blocked_navigation", lambda: None)():
+                raise ValueError(f"navigation was blocked: {blocked}")
+        except ValueError as exc:
+            return f"Error: {exc}"
+        except NotImplementedError as exc:
+            return f"Error: {exc}"
+        except Exception as exc:
+            return f"Error executing computer_use '{action}': {type(exc).__name__}: {exc}"
+
+        # Return a fresh screenshot so the model sees the result of its action.
+        try:
+            png = await backend.screenshot()
+            png = self._downscale_png(png, target)
+        except ImportError as exc:
+            return f"Error: {exc}"
+        except Exception as exc:
+            return f"{status}\n(Could not capture screenshot: {type(exc).__name__}: {exc})"
+
+        label = f"{status} | screen {target[0]}x{target[1]} ({backend.environment})"
+        return build_image_content_blocks(png, "image/png", "", label)
+
+    async def close(self) -> None:
+        await self._backends.close()

nanobot/agent/tools/computer_use_backends/__init__.py

@@ -0,0 +1 @@
+"""Computer-use backend adapters."""

nanobot/agent/tools/computer_use_backends/base.py

@@ -0,0 +1,139 @@
+"""Backend interface for the ``computer_use`` tool.
+
+A backend is the *actuator* + *screenshot source* for one execution environment
+(the local desktop, a headless browser, a VM, ...). The tool layer owns the
+agent loop, coordinate scaling, screenshot downscaling and safety gating; a
+backend only has to perform primitive actions and grab a screenshot.
+
+Coordinate contract: every ``x``/``y`` passed to a backend is already in **real
+device pixels** (the same pixel space as :meth:`screenshot`). The tool scales the
+model's target-space coordinates to real pixels before calling the backend, so
+backends never deal with the downscaled space.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from abc import ABC, abstractmethod
+from collections import OrderedDict
+from collections.abc import Awaitable, Callable
+from typing import Any
+
+from nanobot.agent.tools.context import current_request_session_key
+
+
+class ComputerBackend(ABC):
+    """Primitive GUI actions + screenshot for one execution environment."""
+
+    #: "desktop" or "browser" — surfaced to the model so it knows the context.
+    environment: str = "desktop"
+
+    @abstractmethod
+    async def dimensions(self) -> tuple[int, int]:
+        """Return the real screenshot pixel size as ``(width, height)``."""
+
+    @abstractmethod
+    async def screenshot(self) -> bytes:
+        """Return a PNG screenshot of the current screen at real pixel size."""
+
+    @abstractmethod
+    async def click(self, x: int, y: int, button: str = "left", count: int = 1) -> None:
+        """Click at ``(x, y)``. ``button`` in {left,right,middle}; ``count`` for double/triple."""
+
+    @abstractmethod
+    async def move(self, x: int, y: int) -> None:
+        """Move the cursor to ``(x, y)`` without clicking."""
+
+    @abstractmethod
+    async def drag(self, x: int, y: int) -> None:
+        """Press at the current cursor position and drag to ``(x, y)``, then release."""
+
+    @abstractmethod
+    async def scroll(self, x: int, y: int, direction: str, amount: int) -> None:
+        """Scroll at ``(x, y)``. ``direction`` in {up,down,left,right}; ``amount`` in clicks."""
+
+    @abstractmethod
+    async def type_text(self, text: str) -> None:
+        """Type ``text`` at the current focus."""
+
+    @abstractmethod
+    async def key(self, combo: str) -> None:
+        """Press a key or combo, e.g. ``"ctrl+s"`` / ``"Enter"`` (backend-specific syntax)."""
+
+    async def navigate(self, url: str) -> None:
+        """Navigate to ``url`` (browser backends only)."""
+        raise NotImplementedError(
+            f"'navigate' is not supported by the {self.environment} backend"
+        )
+
+    async def close(self) -> None:
+        """Release any resources (browser process, etc.). Safe to call repeatedly."""
+        return None
+
+
+class SessionBackendPool:
+    """Keep stateful backends isolated by nanobot session."""
+
+    def __init__(
+        self,
+        factory: Callable[[], Any],
+        injected: Any = None,
+        *,
+        max_backends: int = 8,
+        finalizer: Callable[[], Awaitable[None]] | None = None,
+    ) -> None:
+        if max_backends < 1:
+            raise ValueError("max_backends must be at least 1")
+        self._factory = factory
+        self._injected = injected
+        self._max_backends = max_backends
+        self._finalizer = finalizer
+        self._backends: OrderedDict[str, Any] = OrderedDict()
+        self._lock = asyncio.Lock()
+        self._closed = False
+
+    async def get(self) -> Any:
+        async with self._lock:
+            if self._closed:
+                raise RuntimeError("computer-use backend pool is closed")
+            if self._injected is not None:
+                return self._injected
+            key = current_request_session_key() or "default"
+            backend = self._backends.get(key)
+            if backend is not None:
+                self._backends.move_to_end(key)
+                return backend
+            if len(self._backends) >= self._max_backends:
+                _, stale = self._backends.popitem(last=False)
+                await stale.close()
+            backend = self._factory()
+            self._backends[key] = backend
+            return backend
+
+    async def close(self) -> None:
+        async with self._lock:
+            if self._closed:
+                return
+            self._closed = True
+            backends = (
+                [self._injected]
+                if self._injected is not None
+                else list(self._backends.values())
+            )
+            self._injected = None
+            self._backends.clear()
+            finalizer, self._finalizer = self._finalizer, None
+        results = await asyncio.gather(
+            *(backend.close() for backend in backends if backend is not None),
+            return_exceptions=True,
+        )
+        errors = [result for result in results if isinstance(result, BaseException)]
+        if finalizer is not None:
+            try:
+                await finalizer()
+            except BaseException as exc:
+                errors.append(exc)
+        if len(errors) == 1:
+            raise errors[0]
+        if errors:
+            raise BaseExceptionGroup("failed to close computer-use backends", errors)

nanobot/agent/tools/computer_use_backends/browser_playwright.py

@@ -0,0 +1,382 @@
+"""Playwright backend shared by browser and computer_use."""
+
+from __future__ import annotations
+
+import asyncio
+import importlib
+from collections.abc import Sequence
+from typing import Any, cast
+from urllib.parse import urlparse, urlunparse
+
+from loguru import logger
+
+from nanobot.agent.tools.computer_use_backends.base import ComputerBackend
+from nanobot.security.network import validate_url_target
+
+_MISSING = (
+    "Browser computer-use backend needs 'playwright'. Install with: "
+    "pip install 'nanobot-ai[computer-use]' && playwright install chromium"
+)
+
+_SCROLL_PIXELS = 100  # one "scroll click" ~= this many pixels
+
+# Tags visible interactive elements with data-nanobot-ref and returns a compact
+# list. Refs are reassigned per call. Used by DOM/accessibility mode.
+_SNAPSHOT_JS = r"""
+(max) => {
+  const SEL = 'a,button,input,textarea,select,[role=button],[role=link],[role=checkbox],[role=radio],[role=tab],[role=menuitem],[role=switch],[onclick],[contenteditable=""],[contenteditable=true]';
+  const out = [];
+  let ref = 0;
+  for (const el of document.querySelectorAll(SEL)) {
+    const r = el.getBoundingClientRect();
+    const s = getComputedStyle(el);
+    if (r.width <= 0 || r.height <= 0) continue;
+    if (s.visibility === 'hidden' || s.display === 'none' || s.opacity === '0') continue;
+    ref++;
+    el.setAttribute('data-nanobot-ref', String(ref));
+    let name = (el.getAttribute('aria-label') || el.innerText || el.value ||
+                el.getAttribute('placeholder') || el.getAttribute('name') ||
+                el.getAttribute('title') || '');
+    name = name.replace(/\s+/g, ' ').trim().slice(0, 120);
+    out.push({
+      ref: ref,
+      tag: el.tagName.toLowerCase(),
+      role: el.getAttribute('role') || '',
+      type: el.getAttribute('type') || '',
+      name: name,
+      href: el.getAttribute('href') || ''
+    });
+    if (out.length >= max) break;
+  }
+  return out;
+}
+"""
+
+# CUA/xdotool-ish modifier names -> Playwright modifiers.
+_MODIFIERS = {
+    "ctrl": "Control", "control": "Control",
+    "alt": "Alt", "option": "Alt",
+    "shift": "Shift",
+    "cmd": "Meta", "meta": "Meta", "super": "Meta", "win": "Meta",
+}
+# Common single-key names -> Playwright key names.
+_KEYS = {
+    "return": "Enter", "enter": "Enter", "tab": "Tab", "esc": "Escape",
+    "escape": "Escape", "backspace": "Backspace", "delete": "Delete",
+    "space": "Space", "up": "ArrowUp", "down": "ArrowDown",
+    "left": "ArrowLeft", "right": "ArrowRight",
+    "page_down": "PageDown", "pagedown": "PageDown",
+    "page_up": "PageUp", "pageup": "PageUp", "home": "Home", "end": "End",
+}
+
+
+def _validate_browser_url(
+    url: str,
+    allowed_domains: Sequence[str] = (),
+    *,
+    navigation: bool = True,
+) -> tuple[bool, str]:
+    if url == "about:blank":
+        return True, ""
+
+    parsed = urlparse(url)
+    if not navigation and parsed.scheme in {"blob", "data"}:
+        return True, ""
+
+    target = url
+    if parsed.scheme in {"ws", "wss"}:
+        target = urlunparse(parsed._replace(scheme="https" if parsed.scheme == "wss" else "http"))
+
+    if navigation and allowed_domains:
+        host = (parsed.hostname or "").rstrip(".").lower()
+        allowed = any(
+            normalized and (host == normalized or host.endswith(f".{normalized}"))
+            for domain in allowed_domains
+            if (normalized := domain.strip().lstrip(".").rstrip(".").lower())
+        )
+        if not allowed:
+            return False, f"host {host or '<missing>'} is not in allowed_domains"
+
+    return validate_url_target(target)
+
+
+def _playwright_key(combo: str) -> str:
+    parts = [p.strip() for p in combo.split("+") if p.strip()]
+    out: list[str] = []
+    for part in parts:
+        low = part.lower()
+        if low in _MODIFIERS:
+            out.append(_MODIFIERS[low])
+        elif low in _KEYS:
+            out.append(_KEYS[low])
+        elif len(part) == 1:
+            out.append(part)
+        else:
+            out.append(part.capitalize())
+    return "+".join(out)
+
+
+class BrowserRuntime:
+    """One lazily started browser process shared by isolated session contexts."""
+
+    def __init__(self, *, headless: bool = True) -> None:
+        self._headless = headless
+        self._lock = asyncio.Lock()
+        self._playwright: Any = None
+        self._browser: Any = None
+
+    async def get(self) -> Any:
+        if self._browser is not None:
+            return self._browser
+        async with self._lock:
+            if self._browser is not None:
+                return self._browser
+            try:
+                playwright = importlib.import_module("playwright.async_api")
+                async_playwright = cast(Any, playwright).async_playwright
+            except ImportError as exc:
+                raise ImportError(_MISSING) from exc
+            self._playwright = await async_playwright().start()
+            try:
+                self._browser = await self._playwright.chromium.launch(
+                    headless=self._headless
+                )
+            except BaseException:
+                await self.close()
+                raise
+        return self._browser
+
+    async def close(self) -> None:
+        browser, playwright = self._browser, self._playwright
+        self._browser = self._playwright = None
+        errors: list[BaseException] = []
+        closers = (
+            browser.close if browser is not None else None,
+            playwright.stop if playwright is not None else None,
+        )
+        for close in closers:
+            if close is None:
+                continue
+            try:
+                await close()
+            except BaseException as exc:
+                errors.append(exc)
+        if len(errors) == 1:
+            raise errors[0]
+        if errors:
+            raise BaseExceptionGroup("failed to close browser runtime", errors)
+
+
+class BrowserBackend(ComputerBackend):
+    environment = "browser"
+
+    def __init__(
+        self,
+        *,
+        width: int = 1280,
+        height: int = 800,
+        headless: bool = True,
+        start_url: str = "about:blank",
+        allowed_domains: Sequence[str] = (),
+        runtime: BrowserRuntime | None = None,
+    ) -> None:
+        self._width = width
+        self._height = height
+        self._start_url = start_url
+        self._allowed_domains = tuple(allowed_domains)
+        self._runtime = runtime or BrowserRuntime(headless=headless)
+        self._owns_runtime = runtime is None
+        self._context: Any = None
+        self._page: Any = None
+        self._last_pos = (0, 0)
+        self._blocked_navigation: str | None = None
+
+    async def _require_url(self, url: str, label: str) -> None:
+        ok, error = await asyncio.to_thread(
+            _validate_browser_url,
+            url,
+            self._allowed_domains,
+        )
+        if not ok:
+            raise ValueError(f"{label} is blocked: {error}")
+
+    async def _route_request(self, route: Any) -> None:
+        request = route.request
+        navigation = bool(request.is_navigation_request())
+        ok, error = await asyncio.to_thread(
+            _validate_browser_url,
+            request.url,
+            self._allowed_domains,
+            navigation=navigation,
+        )
+        if ok:
+            await route.continue_()
+            return
+        if navigation:
+            self._blocked_navigation = error
+        logger.warning("Blocked browser request to {}: {}", request.url, error)
+        await route.abort("blockedbyclient")
+
+    async def _route_web_socket(self, web_socket: Any) -> None:
+        ok, error = await asyncio.to_thread(
+            _validate_browser_url,
+            web_socket.url,
+            self._allowed_domains,
+            navigation=False,
+        )
+        if not ok:
+            logger.warning("Blocked browser WebSocket to {}: {}", web_socket.url, error)
+            await web_socket.close(code=1008, reason="Blocked by nanobot network policy")
+            return
+        await web_socket.connect_to_server()
+
+    def pop_blocked_navigation(self) -> str | None:
+        error = self._blocked_navigation
+        self._blocked_navigation = None
+        return error
+
+    async def _ensure(self) -> Any:
+        if self._page is not None:
+            return self._page
+        await self._require_url(self._start_url, "start_url")
+        try:
+            browser = await self._runtime.get()
+            self._context = await browser.new_context(
+                viewport={"width": self._width, "height": self._height},
+                device_scale_factor=1,
+                service_workers="block",
+            )
+            await self._context.route("**/*", self._route_request)
+            await self._context.route_web_socket("**/*", self._route_web_socket)
+            self._page = await self._context.new_page()
+            if self._start_url != "about:blank":
+                await self._page.goto(self._start_url)
+            return self._page
+        except BaseException:
+            await self.close()
+            raise
+
+    async def dimensions(self) -> tuple[int, int]:
+        await self._ensure()
+        vp = self._page.viewport_size or {"width": self._width, "height": self._height}
+        return vp["width"], vp["height"]
+
+    async def screenshot(self) -> bytes:
+        page = await self._ensure()
+        return await page.screenshot()
+
+    async def click(self, x: int, y: int, button: str = "left", count: int = 1) -> None:
+        page = await self._ensure()
+        await page.mouse.click(x, y, button=button, click_count=count)
+        self._last_pos = (x, y)
+
+    async def move(self, x: int, y: int) -> None:
+        page = await self._ensure()
+        await page.mouse.move(x, y)
+        self._last_pos = (x, y)
+
+    async def drag(self, x: int, y: int) -> None:
+        page = await self._ensure()
+        sx, sy = self._last_pos
+        await page.mouse.move(sx, sy)
+        await page.mouse.down()
+        await page.mouse.move(x, y)
+        await page.mouse.up()
+        self._last_pos = (x, y)
+
+    async def scroll(self, x: int, y: int, direction: str, amount: int) -> None:
+        page = await self._ensure()
+        await page.mouse.move(x, y)
+        pixels = max(1, amount) * _SCROLL_PIXELS
+        dx = pixels if direction == "right" else -pixels if direction == "left" else 0
+        dy = pixels if direction == "down" else -pixels if direction == "up" else 0
+        await page.mouse.wheel(dx, dy)
+
+    async def type_text(self, text: str) -> None:
+        page = await self._ensure()
+        await page.keyboard.type(text)
+
+    async def key(self, combo: str) -> None:
+        page = await self._ensure()
+        key = _playwright_key(combo)
+        if key:
+            await page.keyboard.press(key)
+
+    async def navigate(self, url: str) -> None:
+        await self._require_url(url, "navigation")
+        page = await self._ensure()
+        await page.goto(url)
+        self._last_pos = (0, 0)
+
+    # --- DOM / accessibility mode (act by element ref, not pixels) ---
+
+    async def dom_snapshot(self, max_elements: int = 200) -> list[dict[str, Any]]:
+        """Tag visible interactive elements with ``data-nanobot-ref`` and return them.
+
+        Each entry: ``{ref, tag, role, type, name, href}``. Refs are reassigned on
+        every snapshot, so callers should act on the latest snapshot.
+        """
+        page = await self._ensure()
+        return cast(list[dict[str, Any]], await page.evaluate(_SNAPSHOT_JS, max_elements))
+
+    def _ref_selector(self, ref: int) -> str:
+        return f'[data-nanobot-ref="{int(ref)}"]'
+
+    async def click_ref(self, ref: int) -> None:
+        page = await self._ensure()
+        await page.click(self._ref_selector(ref), timeout=5000)
+
+    async def fill_ref(self, ref: int, text: str, submit: bool = False) -> None:
+        page = await self._ensure()
+        sel = self._ref_selector(ref)
+        await page.fill(sel, text, timeout=5000)
+        if submit:
+            await page.press(sel, "Enter")
+
+    async def select_ref(self, ref: int, value: str) -> None:
+        page = await self._ensure()
+        sel = self._ref_selector(ref)
+        try:
+            await page.select_option(sel, value, timeout=3000)
+        except Exception:
+            # Models usually pass the visible label, not the option value.
+            await page.select_option(sel, label=value, timeout=3000)
+
+    async def scroll_page(self, direction: str, amount: int) -> None:
+        page = await self._ensure()
+        pixels = max(1, amount) * _SCROLL_PIXELS
+        dx = pixels if direction == "right" else -pixels if direction == "left" else 0
+        dy = pixels if direction == "down" else -pixels if direction == "up" else 0
+        await page.evaluate("([x, y]) => window.scrollBy(x, y)", [dx, dy])
+
+    async def go_back(self) -> None:
+        page = await self._ensure()
+        await page.go_back()
+
+    async def read_text(self, max_chars: int = 4000) -> str:
+        page = await self._ensure()
+        txt = await page.evaluate("() => document.body ? document.body.innerText : ''")
+        return (txt or "")[:max_chars]
+
+    async def current_url(self) -> str:
+        page = await self._ensure()
+        return page.url
+
+    async def close(self) -> None:
+        context = self._context
+        self._context = self._page = None
+        error: BaseException | None = None
+        if context is not None:
+            try:
+                await context.close()
+            except BaseException as exc:
+                error = exc
+        if self._owns_runtime:
+            try:
+                await self._runtime.close()
+            except BaseException as exc:
+                if error is not None:
+                    raise BaseExceptionGroup("failed to close browser backend", [error, exc])
+                raise
+        if error is not None:
+            raise error

nanobot/agent/tools/computer_use_backends/desktop_pyautogui.py

@@ -0,0 +1,134 @@
+"""PyAutoGUI desktop backend with HiDPI coordinate correction."""
+
+from __future__ import annotations
+
+import asyncio
+import io
+from typing import Any
+
+from nanobot.agent.tools.computer_use_backends.base import ComputerBackend
+
+_MISSING = (
+    "Desktop computer-use backend needs 'pyautogui' and 'pillow'. "
+    "Install with: pip install 'nanobot-ai[computer-use]'"
+)
+
+# xdotool/CUA-style key names -> PyAutoGUI key names.
+_KEY_ALIASES = {
+    "return": "enter",
+    "ctrl": "ctrl",
+    "control": "ctrl",
+    "cmd": "command",
+    "super": "win",
+    "win": "win",
+    "page_down": "pagedown",
+    "page_up": "pageup",
+    "pagedown": "pagedown",
+    "pageup": "pageup",
+    "esc": "esc",
+    "escape": "esc",
+}
+
+
+class DesktopBackend(ComputerBackend):
+    environment = "desktop"
+
+    def __init__(self) -> None:
+        self._pg: Any = None
+        self._ratio_x = 1.0
+        self._ratio_y = 1.0
+        self._dims: tuple[int, int] | None = None
+
+    def _ensure(self) -> Any:
+        if self._pg is not None:
+            return self._pg
+        try:
+            import pyautogui  # noqa: PLC0415
+        except Exception as exc:  # ImportError, or platform display errors
+            raise ImportError(_MISSING) from exc
+        self._pg = pyautogui
+        return pyautogui
+
+    def _grab_png_and_size(self) -> tuple[bytes, int, int]:
+        pg = self._ensure()
+        img = pg.screenshot()
+        buf = io.BytesIO()
+        img.save(buf, format="PNG")
+        width, height = img.size
+        # Refresh logical<->physical ratio from the actual grab.
+        try:
+            logical_w, logical_h = pg.size()
+            self._ratio_x = (logical_w / width) if width else 1.0
+            self._ratio_y = (logical_h / height) if height else 1.0
+        except Exception:
+            self._ratio_x = self._ratio_y = 1.0
+        self._dims = (width, height)
+        return buf.getvalue(), width, height
+
+    def _to_logical(self, x: int, y: int) -> tuple[int, int]:
+        return round(x * self._ratio_x), round(y * self._ratio_y)
+
+    async def dimensions(self) -> tuple[int, int]:
+        if self._dims is not None:
+            return self._dims
+        _, w, h = await asyncio.to_thread(self._grab_png_and_size)
+        return w, h
+
+    async def screenshot(self) -> bytes:
+        png, _, _ = await asyncio.to_thread(self._grab_png_and_size)
+        return png
+
+    async def click(self, x: int, y: int, button: str = "left", count: int = 1) -> None:
+        pg = self._ensure()
+        lx, ly = self._to_logical(x, y)
+        await asyncio.to_thread(pg.click, lx, ly, clicks=count, button=button)
+
+    async def move(self, x: int, y: int) -> None:
+        pg = self._ensure()
+        lx, ly = self._to_logical(x, y)
+        await asyncio.to_thread(pg.moveTo, lx, ly)
+
+    async def drag(self, x: int, y: int) -> None:
+        pg = self._ensure()
+        lx, ly = self._to_logical(x, y)
+        await asyncio.to_thread(
+            pg.dragTo,
+            lx,
+            ly,
+            duration=0.3,
+            tween=pg.easeInOutQuad,
+            button="left",
+        )
+
+    async def scroll(self, x: int, y: int, direction: str, amount: int) -> None:
+        pg = self._ensure()
+        lx, ly = self._to_logical(x, y)
+        clicks = max(1, amount)
+        await asyncio.to_thread(pg.moveTo, lx, ly)
+        if direction in ("up", "down"):
+            await asyncio.to_thread(pg.scroll, clicks if direction == "up" else -clicks)
+        else:
+            await asyncio.to_thread(pg.hscroll, clicks if direction == "right" else -clicks)
+
+    async def type_text(self, text: str) -> None:
+        if not text.isascii():
+            raise ValueError(
+                "desktop text input supports ASCII key events only; "
+                "use the browser backend for Unicode text"
+            )
+        pg = self._ensure()
+        await asyncio.to_thread(pg.typewrite, text, 0.01)
+
+    async def key(self, combo: str) -> None:
+        pg = self._ensure()
+        keys = [
+            _KEY_ALIASES.get(part.strip().lower(), part.strip().lower())
+            for part in combo.split("+")
+            if part.strip()
+        ]
+        if not keys:
+            return
+        if len(keys) == 1:
+            await asyncio.to_thread(pg.press, keys[0])
+        else:
+            await asyncio.to_thread(pg.hotkey, *keys)

nanobot/agent/tools/loader.py

@@ -187,5 +187,8 @@ class _LegacyErrorPrefixTool(Tool):
             return ToolResult.error(result)
         return result
 
+    async def close(self) -> None:
+        await self._wrapped.close()
+
     def __getattr__(self, name: str) -> Any:
         return getattr(self._wrapped, name)

nanobot/agent/tools/registry.py

@@ -200,6 +200,19 @@ class ToolRegistry:
         except Exception as e:
             return ToolResult.error(f"Error executing {name}: {str(e)}" + hint)
 
+    async def close(self) -> None:
+        """Close every registered tool, attempting all cleanups."""
+        errors: list[BaseException] = []
+        for tool in self._tools.values():
+            try:
+                await tool.close()
+            except BaseException as exc:
+                errors.append(exc)
+        if len(errors) == 1:
+            raise errors[0]
+        if errors:
+            raise BaseExceptionGroup("failed to close tools", errors)
+
     @property
     def tool_names(self) -> list[str]:
         """Get list of registered tool names."""

nanobot/config/schema.py

@@ -12,7 +12,9 @@ from nanobot.config_base import Base
 from nanobot.cron.types import CronSchedule
 
 if TYPE_CHECKING:
+    from nanobot.agent.tools.browser_tool import BrowserToolConfig
     from nanobot.agent.tools.cli_apps import CliAppsToolConfig
+    from nanobot.agent.tools.computer_use import ComputerUseToolConfig
     from nanobot.agent.tools.filesystem import FileToolsConfig
     from nanobot.agent.tools.image_generation import ImageGenerationToolConfig
     from nanobot.agent.tools.self import MyToolConfig
@@ -399,6 +401,16 @@ class ToolsConfig(Base):
     """
 
     web: WebToolsConfig = Field(default_factory=lambda: _lazy_default("nanobot.agent.tools.web", "WebToolsConfig"))
+    browser: BrowserToolConfig = Field(
+        default_factory=lambda: _lazy_default(
+            "nanobot.agent.tools.browser_tool", "BrowserToolConfig"
+        )
+    )
+    computer_use: ComputerUseToolConfig = Field(
+        default_factory=lambda: _lazy_default(
+            "nanobot.agent.tools.computer_use", "ComputerUseToolConfig"
+        )
+    )
     exec: ExecToolConfig = Field(default_factory=lambda: _lazy_default("nanobot.agent.tools.shell", "ExecToolConfig"))
     file: FileToolsConfig = Field(default_factory=lambda: _lazy_default("nanobot.agent.tools.filesystem", "FileToolsConfig"))
     cli_apps: CliAppsToolConfig = Field(default_factory=lambda: _lazy_default("nanobot.agent.tools.cli_apps", "CliAppsToolConfig"))
@@ -670,7 +682,9 @@ def _resolve_tool_config_refs() -> None:
     """
     import sys
 
+    from nanobot.agent.tools.browser_tool import BrowserToolConfig
     from nanobot.agent.tools.cli_apps import CliAppsToolConfig
+    from nanobot.agent.tools.computer_use import ComputerUseToolConfig
     from nanobot.agent.tools.filesystem import FileToolsConfig
     from nanobot.agent.tools.image_generation import ImageGenerationToolConfig
     from nanobot.agent.tools.self import MyToolConfig
@@ -680,6 +694,8 @@ def _resolve_tool_config_refs() -> None:
     # Re-export into this module's namespace
     mod = sys.modules[__name__]
     mod.ExecToolConfig = ExecToolConfig  # type: ignore[attr-defined]
+    mod.BrowserToolConfig = BrowserToolConfig  # type: ignore[attr-defined]
+    mod.ComputerUseToolConfig = ComputerUseToolConfig  # type: ignore[attr-defined]
     mod.FileToolsConfig = FileToolsConfig  # type: ignore[attr-defined]
     mod.CliAppsToolConfig = CliAppsToolConfig  # type: ignore[attr-defined]
     mod.WebToolsConfig = WebToolsConfig  # type: ignore[attr-defined]

nanobot/providers/openai_compat_provider.py

@@ -671,6 +671,73 @@ class OpenAICompatProvider(LLMProvider):
             dumped = str(content)
         return dumped or "(empty)"
 
+    @classmethod
+    def _move_tool_images_to_user(
+        cls,
+        messages: list[dict[str, Any]],
+    ) -> list[dict[str, Any]]:
+        """Adapt multimodal tool results to Chat Completions' text-only tool role."""
+        updated: list[dict[str, Any]] = []
+        pending_images: list[dict[str, Any]] = []
+
+        def flush_images(next_message: dict[str, Any] | None = None) -> None:
+            if not pending_images:
+                if next_message is not None:
+                    updated.append(next_message)
+                return
+            content: list[dict[str, Any]] = [
+                *pending_images,
+                {"type": "text", "text": "Images returned by the preceding tool call(s)."},
+            ]
+            pending_images.clear()
+            if next_message is not None and next_message.get("role") == "user":
+                existing = next_message.get("content")
+                if isinstance(existing, str):
+                    content.append({"type": "text", "text": existing})
+                elif isinstance(existing, list):
+                    content.extend(cast(list[dict[str, Any]], existing))
+                updated.append({**next_message, "content": content})
+            else:
+                updated.append({"role": "user", "content": content})
+                if next_message is not None:
+                    updated.append(next_message)
+
+        for message in messages:
+            content = message.get("content")
+            if message.get("role") == "tool" and isinstance(content, list):
+                blocks = cast(list[object], content)
+                images: list[dict[str, Any]] = []
+                text_blocks: list[object] = []
+                for block in blocks:
+                    if isinstance(block, dict):
+                        block_data = cast(dict[str, Any], block)
+                        image_url = block_data.get("image_url")
+                        if block_data.get("type") == "image_url" and isinstance(
+                            image_url, dict
+                        ):
+                            images.append({"type": "image_url", "image_url": image_url})
+                            continue
+                        text_blocks.append(block_data)
+                    else:
+                        text_blocks.append(block)
+                if images:
+                    updated.append({
+                        **message,
+                        "content": (
+                            cls._coerce_content_to_string(text_blocks)
+                            if text_blocks
+                            else "(image returned)"
+                        ),
+                    })
+                    pending_images.extend(images)
+                    continue
+            if message.get("role") != "tool":
+                flush_images(message)
+            else:
+                updated.append(message)
+        flush_images()
+        return updated
+
     def _sanitize_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
         """Strip non-standard keys, normalize tool_call IDs."""
         sanitized = LLMProvider._sanitize_request_messages(messages, _ALLOWED_MSG_KEYS)
@@ -824,9 +891,10 @@ class OpenAICompatProvider(LLMProvider):
 
         model_name = self._request_model_name(model_name)
 
+        sanitized_messages = self._sanitize_messages(self._sanitize_empty_content(messages))
         kwargs: dict[str, Any] = {
             "model": model_name,
-            "messages": self._sanitize_messages(self._sanitize_empty_content(messages)),
+            "messages": self._move_tool_images_to_user(sanitized_messages),
         }
 
         # GPT-5 and reasoning models (o1/o3/o4) reject temperature when

nanobot/utils/helpers.py

@@ -356,6 +356,7 @@ _TOOL_RESULT_PREVIEW_CHARS = 1200
 _TOOL_RESULTS_DIR = ".nanobot/tool-results"
 _TOOL_RESULT_RETENTION_SECS = 7 * 24 * 60 * 60
 _TOOL_RESULT_MAX_BUCKETS = 32
+_IMAGE_TOKEN_ESTIMATE = 2048
 _TRUNCATED_SUFFIX = "\n... (truncated)"
 
 
@@ -676,6 +677,7 @@ def _estimate_prompt_tokens_with_source(
     reasoning_content, tool_call_id, name, plus per-message framing overhead.
     """
     parts: list[str] = []
+    image_tokens = 0
     for msg in messages:
         content = msg.get("content")
         if isinstance(content, str):
@@ -687,6 +689,8 @@ def _estimate_prompt_tokens_with_source(
                     text = part.get("text", "")
                     if isinstance(text, str) and text:
                         parts.append(text)
+                elif part is not None and part.get("type") in {"image_url", "input_image"}:
+                    image_tokens += _IMAGE_TOKEN_ESTIMATE
 
         tc = msg.get("tool_calls")
         if tc:
@@ -709,7 +713,7 @@ def _estimate_prompt_tokens_with_source(
             _estimate_tools_tokens(enc, tools, leading_separator=bool(parts)) if tools else 0
         )
         message_tokens = len(enc.encode(message_payload)) if message_payload else 0
-        return message_tokens + tool_tokens + per_message_overhead, "tiktoken"
+        return message_tokens + image_tokens + tool_tokens + per_message_overhead, "tiktoken"
     except Exception:
         tool_payload = (
             ("\n" if message_payload else "") + json.dumps(tools, ensure_ascii=False)
@@ -718,7 +722,7 @@ def _estimate_prompt_tokens_with_source(
         )
         payload = message_payload + tool_payload
         estimated = len(payload.encode("utf-8"))
-        return estimated + per_message_overhead, "heuristic"
+        return estimated + image_tokens + per_message_overhead, "heuristic"
 
 
 def estimate_prompt_tokens(
@@ -734,6 +738,7 @@ def estimate_message_tokens(message: dict[str, Any]) -> int:
     """Estimate prompt tokens contributed by one persisted message."""
     content = message.get("content")
     parts: list[str] = []
+    image_tokens = 0
     if isinstance(content, str):
         parts.append(content)
     elif isinstance(content, list):
@@ -743,6 +748,8 @@ def estimate_message_tokens(message: dict[str, Any]) -> int:
                 text = part.get("text", "")
                 if isinstance(text, str) and text:
                     parts.append(text)
+            elif part is not None and part.get("type") in {"image_url", "input_image"}:
+                image_tokens += _IMAGE_TOKEN_ESTIMATE
             else:
                 parts.append(json.dumps(raw_part, ensure_ascii=False))
     elif content is not None:
@@ -760,13 +767,13 @@ def estimate_message_tokens(message: dict[str, Any]) -> int:
         parts.append(rc)
 
     payload = "\n".join(parts)
-    if not payload:
+    if not payload and not image_tokens:
         return 4
     try:
         enc = _get_token_encoding()
-        return max(4, len(enc.encode(payload)) + 4)
+        return max(4, len(enc.encode(payload)) + image_tokens + 4)
     except Exception:
-        return max(4, len(payload.encode("utf-8")) + 4)
+        return max(4, len(payload.encode("utf-8")) + image_tokens + 4)
 
 
 def estimate_prompt_tokens_chain(

nanobot/webui/settings_api.py

@@ -1261,6 +1261,11 @@ def settings_payload(
                 "use_jina_reader": config.tools.web.fetch.use_jina_reader,
             },
         },
+        "computer_use": {
+            "browser_enabled": config.tools.browser.enable,
+            "enabled": config.tools.computer_use.enable,
+            "backend": config.tools.computer_use.backend,
+        },
         "api": {
             "host": config.api.host,
             "port": config.api.port,
@@ -2045,6 +2050,30 @@ def update_network_safety_settings(query: QueryParams) -> dict[str, Any]:
     return settings_payload(requires_restart=changed)
 
 
+def update_computer_use_settings(query: QueryParams) -> dict[str, Any]:
+    raw_browser = _query_first_alias(query, "browser_enabled", "browserEnabled")
+    raw_computer = _query_first_alias(query, "enabled", "computerEnabled")
+    if raw_browser is None and raw_computer is None:
+        raise WebUISettingsError("browser_enabled or enabled is required")
+
+    config = load_config()
+    changed = False
+    if raw_browser is not None:
+        browser_enabled = _parse_bool(raw_browser, "browser_enabled")
+        if config.tools.browser.enable != browser_enabled:
+            config.tools.browser.enable = browser_enabled
+            changed = True
+    if raw_computer is not None:
+        computer_enabled = _parse_bool(raw_computer, "enabled")
+        if config.tools.computer_use.enable != computer_enabled:
+            config.tools.computer_use.enable = computer_enabled
+            changed = True
+
+    if changed:
+        save_config(config)
+    return settings_payload(requires_restart=changed)
+
+
 def update_web_search_settings(query: QueryParams) -> dict[str, Any]:
     provider_name = (_query_first(query, "provider") or "").strip().lower()
     provider_option = _WEB_SEARCH_PROVIDER_BY_NAME.get(provider_name)

nanobot/webui/settings_routes.py

@@ -64,6 +64,7 @@ from nanobot.webui.settings_api import (
     settings_usage_payload,
     update_agent_settings,
     update_api_settings,
+    update_computer_use_settings,
     update_image_generation_settings,
     update_model_call_order,
     update_model_configuration,
@@ -174,6 +175,8 @@ class WebUISettingsRouter:
             return await self._handle_settings_provider_oauth(request, "logout")
         if path == "/api/settings/web-search/update":
             return self._handle_settings_web_search_update(request)
+        if path == "/api/settings/computer-use/update":
+            return self._handle_settings_computer_use_update(request)
         if path == "/api/settings/api-service":
             return self._handle_settings_api_service(request)
         if path == "/api/settings/api-service/start":
@@ -506,6 +509,15 @@ class WebUISettingsRouter:
             return self._error_response(e.status, e.message)
         return self._json_response(self._with_restart_state(payload, section="browser"))
 
+    def _handle_settings_computer_use_update(self, request: WsRequest) -> Response:
+        if not self._authorized(request):
+            return self._unauthorized()
+        try:
+            payload = update_computer_use_settings(self._query(request))
+        except WebUISettingsError as e:
+            return self._error_response(e.status, e.message)
+        return self._json_response(self._with_restart_state(payload, section="runtime"))
+
     def _handle_settings_api_service(self, request: WsRequest) -> Response:
         if not self._authorized(request):
             return self._unauthorized()

pyproject.toml

@@ -88,6 +88,11 @@ pdf = [
 olostep = [
     "olostep>=0.1.0; python_version < '3.14'",
 ]
+computer-use = [
+    "pyautogui>=0.9.54",
+    "pillow>=10.0.0",
+    "playwright>=1.48.0",
+]
 dev = [
     "pytest>=9.0.0,<10.0.0",
     "pytest-asyncio>=1.3.0,<2.0.0",

tests/agent/test_context_governance.py

@@ -1,6 +1,13 @@
 from nanobot.agent.context_governance import ContextGovernor
 
 
+def _image_result(label: str) -> list[dict]:
+    return [
+        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{label}"}},
+        {"type": "text", "text": label},
+    ]
+
+
 def _assistant_tool_call(call_id: str) -> dict:
     return {
         "role": "assistant",
@@ -37,3 +44,21 @@ def test_drop_orphan_tool_results_drops_duplicate_tool_result() -> None:
     tool_results = [m for m in result if m.get("role") == "tool"]
     assert len(tool_results) == 1
     assert tool_results[0]["content"] == "first"
+
+
+def test_drop_stale_visual_tool_images_keeps_latest_per_tool() -> None:
+    messages = [
+        {"role": "tool", "name": "computer_use", "content": _image_result("history")},
+        {"role": "tool", "name": "computer_use", "content": _image_result("old")},
+        {"role": "tool", "name": "browser", "content": _image_result("browser")},
+        {"role": "tool", "name": "computer_use", "content": _image_result("latest")},
+    ]
+
+    result = ContextGovernor.drop_stale_visual_tool_images(messages, start_index=1)
+
+    assert result is not messages
+    assert result[0]["content"] == messages[0]["content"]
+    assert [block["type"] for block in result[1]["content"]] == ["text", "text"]
+    assert result[2]["content"] == messages[2]["content"]
+    assert result[3]["content"] == messages[3]["content"]
+    assert messages[1]["content"][0]["type"] == "image_url"

tests/agent/test_task_cancel.py

@@ -76,11 +76,13 @@ class TestHandleStop:
 
         loop.subagents.close = close_subagents
         loop._exec_session_manager.close_all = AsyncMock()
+        loop.tools.close = AsyncMock()
         with patch("nanobot.agent.loop.agent_context.close_mcp", AsyncMock()):
             await loop.close_mcp()
 
         assert events == ["turn_cancelled", "resources_closed"]
         assert task.cancelled()
+        loop.tools.close.assert_awaited_once()
 
     @pytest.mark.asyncio
     async def test_close_mcp_serializes_duplicate_cleanup(self):

tests/providers/test_openai_tool_result_images.py

@@ -0,0 +1,50 @@
+from nanobot.providers.openai_compat_provider import OpenAICompatProvider
+
+
+def test_chat_completions_moves_tool_images_after_parallel_results():
+    image = {
+        "type": "image_url",
+        "image_url": {"url": "data:image/png;base64,AAAA"},
+        "_meta": {"path": "screen.png"},
+    }
+    messages = [
+        {"role": "assistant", "content": None, "tool_calls": [{"id": "a"}, {"id": "b"}]},
+        {
+            "role": "tool",
+            "tool_call_id": "a",
+            "content": [image, {"type": "text", "text": "clicked"}],
+        },
+        {"role": "tool", "tool_call_id": "b", "content": "other result"},
+        {"role": "assistant", "content": "done"},
+    ]
+
+    result = OpenAICompatProvider._move_tool_images_to_user(messages)
+
+    assert result[1]["content"] == "clicked"
+    assert result[2] == messages[2]
+    assert result[3]["role"] == "user"
+    assert result[3]["content"][0] == {
+        "type": "image_url",
+        "image_url": {"url": "data:image/png;base64,AAAA"},
+    }
+    assert result[4] == messages[3]
+
+
+def test_chat_completions_merges_tool_images_into_following_user_message():
+    messages = [
+        {
+            "role": "tool",
+            "tool_call_id": "a",
+            "content": [
+                {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}},
+            ],
+        },
+        {"role": "user", "content": "continue"},
+    ]
+
+    result = OpenAICompatProvider._move_tool_images_to_user(messages)
+
+    assert len(result) == 2
+    assert result[0]["content"] == "(image returned)"
+    assert result[1]["role"] == "user"
+    assert result[1]["content"][-1] == {"type": "text", "text": "continue"}

tests/tools/test_browser_tool.py

@@ -0,0 +1,316 @@
+"""Tests for DOM-based browser control."""
+
+from __future__ import annotations
+
+import asyncio
+import io
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from nanobot.agent.tools.browser_tool import BrowserTool, BrowserToolConfig
+from nanobot.agent.tools.computer_use_backends import browser_playwright
+from nanobot.agent.tools.computer_use_backends.browser_playwright import BrowserBackend
+from nanobot.agent.tools.context import RequestContext, request_context
+
+
+class _FakeDomBackend:
+    environment = "browser"
+
+    def __init__(self):
+        self.calls: list[tuple] = []
+        self.elements = [
+            {"ref": 1, "tag": "button", "role": "", "type": "", "name": "Submit", "href": ""},
+            {"ref": 2, "tag": "input", "role": "", "type": "text", "name": "your name", "href": ""},
+        ]
+
+    async def navigate(self, url):
+        self.calls.append(("navigate", url))
+
+    async def dom_snapshot(self, max_elements=200):
+        return self.elements
+
+    async def click_ref(self, ref):
+        self.calls.append(("click", ref))
+
+    async def fill_ref(self, ref, text, submit=False):
+        self.calls.append(("fill", ref, text, submit))
+
+    async def select_ref(self, ref, value):
+        self.calls.append(("select", ref, value))
+
+    async def scroll_page(self, direction, amount):
+        self.calls.append(("scroll", direction, amount))
+
+    async def key(self, combo):
+        self.calls.append(("key", combo))
+
+    async def go_back(self):
+        self.calls.append(("back",))
+
+    async def read_text(self, max_chars=4000):
+        return "the number is 42"
+
+    async def current_url(self):
+        return "http://test.local/page"
+
+    async def screenshot(self):
+        from PIL import Image
+        buf = io.BytesIO()
+        Image.new("RGB", (1280, 800), (0, 0, 0)).save(buf, format="PNG")
+        return buf.getvalue()
+
+    async def close(self):
+        self.calls.append(("close",))
+
+
+def _tool(**kw):
+    fb = _FakeDomBackend()
+    return BrowserTool(BrowserToolConfig(**kw), backend_impl=fb), fb
+
+
+def _route(url: str, *, navigation: bool):
+    return SimpleNamespace(
+        request=SimpleNamespace(
+            url=url,
+            is_navigation_request=MagicMock(return_value=navigation),
+        ),
+        abort=AsyncMock(),
+        continue_=AsyncMock(),
+    )
+
+
+class TestConfigAndMetadata:
+    def test_defaults_off(self):
+        cfg = BrowserToolConfig()
+        assert cfg.enable is False
+        assert cfg.headless is True
+        assert cfg.include_screenshot is False
+        assert cfg.max_elements == 200
+        assert cfg.max_sessions == 8
+
+    def test_enabled_reads_config(self):
+        ctx = MagicMock()
+        ctx.config.browser.enable = True
+        assert BrowserTool.enabled(ctx) is True
+        ctx.config.browser.enable = False
+        assert BrowserTool.enabled(ctx) is False
+
+    def test_create_from_ctx(self):
+        ctx = MagicMock()
+        ctx.config.browser = BrowserToolConfig(enable=True, allowed_domains=["example.com"])
+        tool = BrowserTool.create(ctx)
+        assert isinstance(tool, BrowserTool)
+        assert tool.config.allowed_domains == ["example.com"]
+
+    def test_metadata(self):
+        tool, _ = _tool()
+        assert tool.name == "browser"
+        assert tool.exclusive is True
+        assert tool.read_only is False
+        assert "subagent" not in tool._scopes
+
+    def test_schema_actions(self):
+        tool, _ = _tool()
+        enum = tool.parameters["properties"]["action"]["enum"]
+        for a in ("navigate", "snapshot", "click", "type", "read_text"):
+            assert a in enum
+
+
+class TestDispatch:
+    @pytest.mark.asyncio
+    async def test_navigate_returns_snapshot(self):
+        tool, fb = _tool()
+        result = await tool.execute(action="navigate", url="https://example.com")
+        assert ("navigate", "https://example.com") in fb.calls
+        assert isinstance(result, str)
+        assert "Navigated to https://example.com" in result
+        # snapshot of interactive elements is appended
+        assert '[1] button "Submit"' in result
+        assert '[2] input[text] "your name"' in result
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("action", "kwargs", "expected"),
+        [
+            ("click", {"ref": 1}, ("click", 1)),
+            ("type", {"ref": 2, "text": "Ada", "submit": True}, ("fill", 2, "Ada", True)),
+            ("select", {"ref": 2, "value": "opt1"}, ("select", 2, "opt1")),
+        ],
+    )
+    @pytest.mark.asyncio
+    async def test_element_actions(self, action, kwargs, expected):
+        tool, fb = _tool()
+        result = await tool.execute(action=action, **kwargs)
+        assert expected in fb.calls
+        assert "Interactive elements" in result
+
+    @pytest.mark.asyncio
+    async def test_scroll_and_key_and_back(self):
+        tool, fb = _tool()
+        await tool.execute(action="scroll", scroll_direction="down", scroll_amount=4)
+        await tool.execute(action="key", text="Enter")
+        await tool.execute(action="back")
+        assert ("scroll", "down", 4) in fb.calls
+        assert ("key", "Enter") in fb.calls
+        assert ("back",) in fb.calls
+
+    @pytest.mark.asyncio
+    async def test_read_text_returns_text_no_snapshot(self):
+        tool, _ = _tool()
+        result = await tool.execute(action="read_text")
+        assert isinstance(result, str)
+        assert "the number is 42" in result
+        assert "Interactive elements" not in result
+
+    @pytest.mark.asyncio
+    async def test_include_screenshot_returns_blocks(self):
+        tool, _ = _tool(include_screenshot=True)
+        result = await tool.execute(action="click", ref=1)
+        assert isinstance(result, list)
+        imgs = [b for b in result if b.get("type") == "image_url"]
+        texts = [b for b in result if b.get("type") == "text"]
+        assert imgs and texts
+        assert "Clicked element [1]" in texts[-1]["text"]
+
+    @pytest.mark.asyncio
+    async def test_calls_are_serialized_across_sessions(self):
+        class SlowBackend(_FakeDomBackend):
+            active = 0
+            max_active = 0
+
+            async def dom_snapshot(self, max_elements=200):
+                self.active += 1
+                self.max_active = max(self.max_active, self.active)
+                await asyncio.sleep(0.01)
+                self.active -= 1
+                return await super().dom_snapshot(max_elements)
+
+        backend = SlowBackend()
+        tool = BrowserTool(backend_impl=backend)
+
+        async def snapshot(session: str):
+            with request_context(
+                RequestContext(channel="test", chat_id=session, session_key=session)
+            ):
+                return await tool.execute(action="snapshot")
+
+        await asyncio.gather(snapshot("a"), snapshot("b"))
+
+        assert backend.max_active == 1
+
+
+class TestErrorsAndPolicy:
+    @pytest.mark.parametrize(
+        ("kwargs", "error"),
+        [
+            ({"action": "teleport"}, "unknown action"),
+            ({"action": "click"}, "requires an element 'ref'"),
+        ],
+    )
+    @pytest.mark.asyncio
+    async def test_tool_errors_are_returned_to_model(self, kwargs, error):
+        tool, _ = _tool()
+        result = await tool.execute(**kwargs)
+        assert isinstance(result, str) and error in result
+
+    @pytest.mark.asyncio
+    async def test_backend_blocks_disallowed_navigation(self):
+        backend = BrowserBackend(allowed_domains=["example.com"])
+        page = SimpleNamespace(goto=AsyncMock())
+        backend._page = page
+
+        with pytest.raises(ValueError, match="allowed_domains"):
+            await backend.navigate("https://evil.test/")
+
+        page.goto.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_backend_allows_subdomain_navigation(self, monkeypatch: pytest.MonkeyPatch):
+        check = MagicMock(return_value=(True, ""))
+        monkeypatch.setattr(browser_playwright, "validate_url_target", check)
+        backend = BrowserBackend(allowed_domains=["example.com"])
+        page = SimpleNamespace(goto=AsyncMock())
+        backend._page = page
+
+        await backend.navigate("https://app.example.com/x")
+
+        page.goto.assert_awaited_once_with("https://app.example.com/x")
+        check.assert_called_once()
+
+    @pytest.mark.parametrize(
+        "url",
+        [
+            "file:///etc/passwd",
+            "http://127.0.0.1/",
+            "http://169.254.169.254/latest/meta-data/",
+            "ws://localhost/socket",
+        ],
+    )
+    @pytest.mark.asyncio
+    async def test_browser_network_policy_blocks_local_targets(self, url: str):
+        backend = BrowserBackend()
+        with pytest.raises(ValueError, match="blocked"):
+            await backend.navigate(url)
+
+    @pytest.mark.asyncio
+    async def test_backend_intercepts_blocked_navigation(self):
+        backend = BrowserBackend(allowed_domains=["example.com"])
+        route = _route("https://evil.test/", navigation=True)
+
+        await backend._route_request(route)
+
+        route.abort.assert_awaited_once_with("blockedbyclient")
+        route.continue_.assert_not_awaited()
+        assert "allowed_domains" in (backend.pop_blocked_navigation() or "")
+
+    @pytest.mark.asyncio
+    async def test_backend_intercepts_private_subresource(self):
+        backend = BrowserBackend()
+        route = _route(
+            "http://169.254.169.254/latest/meta-data/",
+            navigation=False,
+        )
+
+        await backend._route_request(route)
+
+        route.abort.assert_awaited_once_with("blockedbyclient")
+        assert backend.pop_blocked_navigation() is None
+
+    @pytest.mark.asyncio
+    async def test_backend_does_not_apply_navigation_allowlist_to_subresources(
+        self, monkeypatch: pytest.MonkeyPatch
+    ):
+        monkeypatch.setattr(
+            browser_playwright,
+            "validate_url_target",
+            MagicMock(return_value=(True, "")),
+        )
+        backend = BrowserBackend(allowed_domains=["example.com"])
+        route = _route("https://cdn.other.test/app.js", navigation=False)
+
+        await backend._route_request(route)
+
+        route.continue_.assert_awaited_once()
+        route.abort.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_backend_intercepts_private_websocket(self):
+        backend = BrowserBackend()
+        web_socket = SimpleNamespace(
+            url="ws://127.0.0.1/socket",
+            close=AsyncMock(),
+            connect_to_server=AsyncMock(),
+        )
+
+        await backend._route_web_socket(web_socket)
+
+        web_socket.close.assert_awaited_once()
+        web_socket.connect_to_server.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_backend_rejects_file_start_url_before_launch(self):
+        backend = BrowserBackend(start_url="file:///etc/passwd")
+        with pytest.raises(ValueError, match="start_url is blocked"):
+            await backend.dimensions()

tests/tools/test_computer_use_tool.py

@@ -0,0 +1,330 @@
+"""Tests for screenshot-based computer control."""
+
+from __future__ import annotations
+
+import asyncio
+import io
+import sys
+from types import SimpleNamespace
+from unittest.mock import MagicMock
+
+import pytest
+
+from nanobot.agent.tools.computer_use import ComputerUseTool, ComputerUseToolConfig
+from nanobot.agent.tools.computer_use_backends.base import ComputerBackend, SessionBackendPool
+from nanobot.agent.tools.computer_use_backends.desktop_pyautogui import DesktopBackend
+from nanobot.agent.tools.context import RequestContext, request_context
+from nanobot.config.schema import ToolsConfig
+
+
+class _FakeBackend(ComputerBackend):
+    """Records actuation calls and serves a solid-colour PNG of a fixed size."""
+
+    environment = "desktop"
+
+    def __init__(self, width: int = 2560, height: int = 1600):
+        self.calls: list[tuple] = []
+        self._w, self._h = width, height
+        self.closed = False
+
+    async def dimensions(self) -> tuple[int, int]:
+        return (self._w, self._h)
+
+    async def screenshot(self) -> bytes:
+        from PIL import Image
+        img = Image.new("RGB", (self._w, self._h), (10, 20, 30))
+        buf = io.BytesIO()
+        img.save(buf, format="PNG")
+        return buf.getvalue()
+
+    async def click(self, x, y, button="left", count=1):
+        self.calls.append(("click", x, y, button, count))
+
+    async def move(self, x, y):
+        self.calls.append(("move", x, y))
+
+    async def drag(self, x, y):
+        self.calls.append(("drag", x, y))
+
+    async def scroll(self, x, y, direction, amount):
+        self.calls.append(("scroll", x, y, direction, amount))
+
+    async def type_text(self, text):
+        self.calls.append(("type", text))
+
+    async def key(self, combo):
+        self.calls.append(("key", combo))
+
+    async def close(self):
+        self.closed = True
+    # navigate() inherited -> raises NotImplementedError (desktop has no navigate)
+
+
+def _split(result):
+    assert isinstance(result, list), f"expected content blocks, got {result!r}"
+    images = [b for b in result if isinstance(b, dict) and b.get("type") == "image_url"]
+    texts = [b for b in result if isinstance(b, dict) and b.get("type") == "text"]
+    return images, texts
+
+
+def _tool(**kw):
+    fb = _FakeBackend(width=kw.pop("w", 2560), height=kw.pop("h", 1600))
+    config = ComputerUseToolConfig(target_width=1280, target_height=800, **kw)
+    tool = ComputerUseTool(config, backend_impl=fb)
+    return tool, fb
+
+
+# --------------------------- config + metadata ---------------------------
+
+class TestConfigAndMetadata:
+    def test_defaults_off(self):
+        cfg = ComputerUseToolConfig()
+        assert cfg.enable is False
+        assert cfg.backend == "desktop"
+        assert (cfg.target_width, cfg.target_height) == (1280, 800)
+        assert cfg.max_sessions == 8
+        assert "require_approval" not in type(cfg).model_fields
+
+    def test_tools_config_accepts_camel_case(self):
+        cfg = ToolsConfig.model_validate({
+            "browser": {"enable": True, "maxSessions": 4},
+            "computerUse": {"enable": True, "backend": "browser", "maxSessions": 6},
+        })
+
+        assert cfg.browser.enable is True
+        assert cfg.browser.max_sessions == 4
+        assert cfg.computer_use.enable is True
+        assert cfg.computer_use.backend == "browser"
+        assert cfg.computer_use.max_sessions == 6
+        dumped = cfg.model_dump(by_alias=True)
+        assert "computerUse" in dumped
+        assert dumped["computerUse"]["maxSessions"] == 6
+
+    def test_enabled_reads_config(self):
+        ctx = MagicMock()
+        ctx.config.computer_use.enable = True
+        assert ComputerUseTool.enabled(ctx) is True
+        ctx.config.computer_use.enable = False
+        assert ComputerUseTool.enabled(ctx) is False
+
+    def test_create_from_ctx(self):
+        ctx = MagicMock()
+        ctx.config.computer_use = ComputerUseToolConfig(
+            enable=True, backend="browser", target_width=1024, target_height=768
+        )
+        tool = ComputerUseTool.create(ctx)
+        assert isinstance(tool, ComputerUseTool)
+        assert tool.config.backend == "browser"
+        assert (tool.config.target_width, tool.config.target_height) == (1024, 768)
+
+    def test_tool_metadata(self):
+        tool, _ = _tool()
+        assert tool.name == "computer_use"
+        assert tool.exclusive is True
+        assert tool.read_only is False
+        assert tool.concurrency_safe is False
+        # not exposed to subagents
+        assert "subagent" not in tool._scopes
+
+    def test_schema_has_action_enum(self):
+        tool, _ = _tool()
+        action = tool.parameters["properties"]["action"]
+        assert "screenshot" in action["enum"]
+        assert "left_click" in action["enum"]
+        assert tool.parameters["required"] == ["action"]
+
+
+# --------------------------- execute dispatch ---------------------------
+
+class TestExecute:
+    @pytest.mark.asyncio
+    async def test_screenshot_returns_image_blocks(self):
+        tool, fb = _tool()
+        result = await tool.execute(action="screenshot")
+        images, texts = _split(result)
+        assert len(images) == 1
+        assert images[0]["image_url"]["url"].startswith("data:image/png;base64,")
+        assert "1280x800" in texts[-1]["text"]
+        assert fb.calls == []  # screenshot performs no actuation
+
+    @pytest.mark.asyncio
+    async def test_left_click_scales_coordinates(self):
+        tool, fb = _tool()  # real 2560x1600 -> target 1280x800 (2x)
+        result = await tool.execute(action="left_click", x=100, y=50)
+        assert fb.calls == [("click", 200, 100, "left", 1)]
+        _, texts = _split(result)
+        assert "left_click at (200, 100)" in texts[-1]["text"]
+
+    @pytest.mark.asyncio
+    async def test_click_clamps_coordinates_to_screen(self):
+        tool, fb = _tool()
+        await tool.execute(action="left_click", x=5000, y=-10)
+        assert fb.calls == [("click", 2559, 0, "left", 1)]
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        ("action", "kwargs", "expected"),
+        [
+            ("double_click", {"x": 10, "y": 10}, ("click", 20, 20, "left", 2)),
+            ("triple_click", {"x": 10, "y": 10}, ("click", 20, 20, "left", 3)),
+            ("right_click", {"x": 5, "y": 5}, ("click", 10, 10, "right", 1)),
+            ("middle_click", {"x": 5, "y": 5}, ("click", 10, 10, "middle", 1)),
+            (
+                "scroll",
+                {"x": 100, "y": 100, "scroll_direction": "down", "scroll_amount": 5},
+                ("scroll", 200, 200, "down", 5),
+            ),
+            ("type", {"text": "hello"}, ("type", "hello")),
+            ("key", {"text": "ctrl+s"}, ("key", "ctrl+s")),
+            ("mouse_move", {"x": 10, "y": 10}, ("move", 20, 20)),
+            ("left_click_drag", {"x": 20, "y": 30}, ("drag", 40, 60)),
+        ],
+    )
+    @pytest.mark.asyncio
+    async def test_actions_dispatch_to_backend(self, action, kwargs, expected):
+        tool, fb = _tool()
+        await tool.execute(action=action, **kwargs)
+        assert fb.calls == [expected]
+
+    @pytest.mark.asyncio
+    async def test_wait(self):
+        tool, fb = _tool()
+        result = await tool.execute(action="wait", duration=0.0)
+        _, texts = _split(result)
+        assert "Waited" in texts[-1]["text"]
+
+    @pytest.mark.parametrize(
+        ("kwargs", "error"),
+        [
+            ({"action": "frobnicate"}, "unknown action"),
+            ({"action": "left_click"}, "requires"),
+            ({"action": "navigate", "url": "https://example.com"}, "Error"),
+        ],
+    )
+    @pytest.mark.asyncio
+    async def test_errors_are_returned_to_model(self, kwargs, error):
+        tool, _ = _tool()
+        result = await tool.execute(**kwargs)
+        assert isinstance(result, str) and error in result
+
+
+@pytest.mark.asyncio
+async def test_backend_pool_isolates_sessions_and_closes_all():
+    created: list[_FakeBackend] = []
+    finalized: list[bool] = []
+
+    def factory():
+        backend = _FakeBackend()
+        created.append(backend)
+        return backend
+
+    async def finalize():
+        finalized.append(all(backend.closed for backend in created))
+
+    pool = SessionBackendPool(factory, finalizer=finalize)
+    with request_context(RequestContext(channel="test", chat_id="a", session_key="test:a")):
+        first = await pool.get()
+        assert await pool.get() is first
+    with request_context(RequestContext(channel="test", chat_id="b", session_key="test:b")):
+        second = await pool.get()
+
+    assert first is not second
+    await pool.close()
+    assert len(created) == 2
+    assert all(backend.closed for backend in created)
+    assert finalized == [True]
+
+    await pool.close()
+    assert finalized == [True]
+
+
+@pytest.mark.asyncio
+async def test_backend_pool_evicts_least_recently_used_session():
+    created: list[_FakeBackend] = []
+
+    def factory():
+        backend = _FakeBackend()
+        created.append(backend)
+        return backend
+
+    pool = SessionBackendPool(factory, max_backends=2)
+    contexts = [
+        RequestContext(channel="test", chat_id=key, session_key=f"test:{key}")
+        for key in ("a", "b", "c")
+    ]
+    with request_context(contexts[0]):
+        first = await pool.get()
+    with request_context(contexts[1]):
+        second = await pool.get()
+    with request_context(contexts[0]):
+        assert await pool.get() is first
+    with request_context(contexts[2]):
+        await pool.get()
+
+    assert first.closed is False
+    assert second.closed is True
+    await pool.close()
+
+
+@pytest.mark.asyncio
+async def test_desktop_tool_serializes_calls_across_sessions():
+    class SlowBackend(_FakeBackend):
+        active = 0
+        max_active = 0
+
+        async def dimensions(self):
+            self.active += 1
+            self.max_active = max(self.max_active, self.active)
+            await asyncio.sleep(0.01)
+            self.active -= 1
+            return await super().dimensions()
+
+    backend = SlowBackend(width=1280, height=800)
+    tool = ComputerUseTool(backend_impl=backend)
+
+    async def screenshot(session: str):
+        with request_context(RequestContext(channel="test", chat_id=session, session_key=session)):
+            return await tool.execute(action="screenshot")
+
+    await asyncio.gather(screenshot("a"), screenshot("b"))
+
+    assert backend.max_active == 1
+
+
+@pytest.mark.asyncio
+async def test_desktop_backend_uses_safe_pyautogui_calls():
+    pg = MagicMock()
+    pg.easeInOutQuad = object()
+    backend = DesktopBackend()
+    backend._pg = pg
+
+    await backend.drag(10, 20)
+    await backend.scroll(10, 20, "down", 3)
+
+    assert pg.dragTo.call_args.kwargs == {
+        "duration": 0.3,
+        "tween": pg.easeInOutQuad,
+        "button": "left",
+    }
+    pg.scroll.assert_called_once_with(-3)
+
+
+@pytest.mark.asyncio
+async def test_desktop_backend_rejects_unicode_instead_of_typing_incorrect_keys():
+    pg = MagicMock()
+    backend = DesktopBackend()
+    backend._pg = pg
+
+    with pytest.raises(ValueError, match="ASCII"):
+        await backend.type_text("你好")
+
+    pg.typewrite.assert_not_called()
+
+
+def test_desktop_backend_preserves_pyautogui_failsafe(monkeypatch):
+    pg = SimpleNamespace(FAILSAFE=True)
+    monkeypatch.setitem(sys.modules, "pyautogui", pg)
+
+    backend = DesktopBackend()
+    assert backend._ensure() is pg
+    assert pg.FAILSAFE is True

tests/tools/test_tool_registry.py

@@ -28,6 +28,18 @@ class _FakeTool(Tool):
     async def execute(self, **kwargs: Any) -> Any:
         return kwargs
 
+
+class _ClosableTool(_FakeTool):
+    def __init__(self, name: str, *, error: BaseException | None = None):
+        super().__init__(name)
+        self.closed = False
+        self.error = error
+
+    async def close(self) -> None:
+        self.closed = True
+        if self.error is not None:
+            raise self.error
+
 def _tool_names(definitions: list[dict[str, Any]]) -> list[str]:
     names: list[str] = []
     for definition in definitions:
@@ -58,6 +70,24 @@ def test_get_definitions_orders_builtins_then_mcp_tools() -> None:
     ]
 
 
+async def test_close_attempts_every_registered_tool() -> None:
+    registry = ToolRegistry()
+    broken = _ClosableTool("broken", error=RuntimeError("close failed"))
+    healthy = _ClosableTool("healthy")
+    registry.register(broken)
+    registry.register(healthy)
+
+    try:
+        await registry.close()
+    except RuntimeError as exc:
+        assert str(exc) == "close failed"
+    else:
+        raise AssertionError("expected close failure")
+
+    assert broken.closed is True
+    assert healthy.closed is True
+
+
 def test_prepare_call_rejects_near_miss_tool_name_with_suggestion() -> None:
     registry = ToolRegistry()
     registry.register(_FakeTool("read_file"))

tests/utils/test_token_estimation.py

@@ -29,6 +29,36 @@ def test_estimate_prompt_tokens_chain_falls_back_without_provider_counter() -> N
     assert source == "tiktoken"
 
 
+def test_image_blocks_have_bounded_token_cost() -> None:
+    text = [{"role": "tool", "content": [{"type": "text", "text": "screen"}]}]
+    small_image = [{
+        "role": "tool",
+        "content": [
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,A"}},
+            {"type": "text", "text": "screen"},
+        ],
+    }]
+    large_image = [{
+        "role": "tool",
+        "content": [
+            {
+                "type": "image_url",
+                "image_url": {"url": "data:image/png;base64," + "A" * 100_000},
+            },
+            {"type": "text", "text": "screen"},
+        ],
+    }]
+
+    text_tokens = estimate_prompt_tokens(text)
+    small_tokens = estimate_prompt_tokens(small_image)
+    large_tokens = estimate_prompt_tokens(large_image)
+
+    assert small_tokens >= text_tokens + 2_000
+    assert large_tokens == small_tokens
+    assert estimate_message_tokens(large_image[0]) >= text_tokens + 2_000
+    assert estimate_message_tokens({"role": "user", "content": small_image[0]["content"][:1]}) > 2_000
+
+
 def test_estimate_prompt_tokens_chain_falls_back_when_provider_counter_fails() -> None:
     tokens, source = estimate_prompt_tokens_chain(
         _BrokenCounterProvider(),

tests/webui/test_settings_api.py

@@ -29,6 +29,7 @@ from nanobot.webui.settings_api import (
     settings_usage_payload,
     update_agent_settings,
     update_api_settings,
+    update_computer_use_settings,
     update_model_call_order,
     update_model_configuration,
     update_network_safety_settings,
@@ -1007,6 +1008,27 @@ def test_settings_payload_includes_network_safety_fields(
     assert payload["advanced"]["ssrf_whitelist_count"] == 1
 
 
+def test_settings_payload_includes_computer_use_tools(
+    tmp_path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    config_path = tmp_path / "config.json"
+    config = Config()
+    config.tools.browser.enable = True
+    config.tools.computer_use.enable = True
+    config.tools.computer_use.backend = "browser"
+    save_config(config, config_path)
+    monkeypatch.setattr("nanobot.config.loader._current_config_path", config_path)
+
+    payload = settings_payload()
+
+    assert payload["computer_use"] == {
+        "browser_enabled": True,
+        "enabled": True,
+        "backend": "browser",
+    }
+
+
 def test_settings_payload_includes_exec_path_flags(
     tmp_path,
     monkeypatch: pytest.MonkeyPatch,
@@ -1374,6 +1396,32 @@ def test_update_network_safety_settings_writes_local_service_flag(
     assert payload["requires_restart"] is True
 
 
+def test_update_computer_use_settings_writes_only_requested_switches(
+    tmp_path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    config_path = tmp_path / "config.json"
+    save_config(Config(), config_path)
+    monkeypatch.setattr("nanobot.config.loader._current_config_path", config_path)
+
+    browser_payload = update_computer_use_settings({"browser_enabled": ["true"]})
+    saved = load_config(config_path)
+    assert saved.tools.browser.enable is True
+    assert saved.tools.computer_use.enable is False
+    assert browser_payload["requires_restart"] is True
+
+    computer_payload = update_computer_use_settings({"computerEnabled": ["true"]})
+    saved = load_config(config_path)
+    assert saved.tools.browser.enable is True
+    assert saved.tools.computer_use.enable is True
+    assert computer_payload["computer_use"]["enabled"] is True
+
+
+def test_update_computer_use_settings_requires_a_switch() -> None:
+    with pytest.raises(WebUISettingsError, match="browser_enabled or enabled"):
+        update_computer_use_settings({})
+
+
 def test_update_network_safety_settings_accepts_legacy_restricted_default_access(
     tmp_path,
     monkeypatch: pytest.MonkeyPatch,

tests/webui/test_settings_routes.py

@@ -138,3 +138,28 @@ async def test_model_preset_mutation_routes(
     assert response.status_code == 200
     assert json.loads(response.body)["routed"] == function_name
     assert captured["query"] == expected_query
+
+
+@pytest.mark.asyncio
+async def test_computer_use_update_route(monkeypatch) -> None:
+    captured: dict[str, object] = {}
+
+    def update(query):
+        captured["query"] = query
+        return {"requires_restart": True}
+
+    monkeypatch.setattr("nanobot.webui.settings_routes.update_computer_use_settings", update)
+    request = SimpleNamespace(
+        path="/api/settings/computer-use/update?browser_enabled=true",
+        headers=Headers(),
+    )
+
+    response = await _router().dispatch(
+        None,
+        request,
+        "/api/settings/computer-use/update",
+    )
+
+    assert response is not None
+    assert response.status_code == 200
+    assert captured["query"] == {"browser_enabled": ["true"]}

webui/src/components/settings/SettingsView.tsx

@@ -139,6 +139,7 @@ import {
   startApiService,
   stopApiService,
   updateAutomation,
+  updateComputerUseSettings,
   updateImageGenerationSettings,
   updateMcpServerTools,
   updateModelCallOrder,
@@ -179,6 +180,7 @@ import type {
   AutomationUpdatePayload,
   CliAppInfo,
   CliAppsPayload,
+  ComputerUseSettingsUpdate,
   ImageGenerationSettingsUpdate,
   McpPresetInfo,
   McpPresetsPayload,
@@ -762,6 +764,7 @@ export function SettingsView({
   const [imageGenerationSaving, setImageGenerationSaving] = useState(false);
   const [transcriptionSaving, setTranscriptionSaving] = useState(false);
   const [networkSafetySaving, setNetworkSafetySaving] = useState(false);
+  const [computerUseSaving, setComputerUseSaving] = useState<"browser" | "computer" | null>(null);
   const [apiService, setApiService] = useState<ApiServicePayload | null>(null);
   const [apiServiceLoading, setApiServiceLoading] = useState(false);
   const [apiServiceAction, setApiServiceAction] = useState<"start" | "stop" | null>(null);
@@ -1002,7 +1005,7 @@ export function SettingsView({
   useEffect(() => {
     if (
       !pageVisible
-      || !["channels", "models", "browser", "runtime"].includes(activeSection)
+      || !["channels", "models", "browser", "runtime", "advanced"].includes(activeSection)
     ) {
       return;
     }
@@ -1559,6 +1562,34 @@ export function SettingsView({
     }
   };
 
+  const setComputerUseEnabled = async (
+    target: "browser" | "computer",
+    enabled: boolean,
+  ) => {
+    if (!settings || computerUseSaving) return;
+    setComputerUseSaving(target);
+    try {
+      if (enabled && !(await installCapabilities(["computer-use"]))) return;
+      const update: ComputerUseSettingsUpdate = target === "browser"
+        ? { browserEnabled: enabled }
+        : { computerEnabled: enabled };
+      const payload = await updateComputerUseSettings(token, update);
+      applyPayload(payload);
+      if (payload.requires_restart) {
+        setPendingRestartSections((prev) => ({
+          ...prev,
+          [target === "browser" ? "browser" : "runtime"]: true,
+        }));
+      }
+      await maybeRestartHostEngine(payload);
+      setError(null);
+    } catch (err) {
+      setError((err as Error).message);
+    } finally {
+      setComputerUseSaving(null);
+    }
+  };
+
   const handleApiServiceAction = async (
     action: "start" | "stop",
     values?: { host: string; port: number; timeout: number; apiKey?: string },
@@ -2238,6 +2269,11 @@ export function SettingsView({
             olostepFeature={featureCatalog.find((feature) => feature.name === "olostep")}
             olostepInstalling={nanobotFeatureAction === "enable:olostep"}
             capabilityError={nanobotFeaturesError}
+            browserAutomationEnabled={settings.computer_use?.browser_enabled ?? false}
+            computerUseFeature={featureCatalog.find((feature) => feature.name === "computer-use")}
+            computerUseSaving={computerUseSaving === "browser"}
+            computerUseInstalling={nanobotFeatureAction === "enable:computer-use"}
+            onToggleBrowserAutomation={(enabled) => void setComputerUseEnabled("browser", enabled)}
           />
         );
       case "channels":
@@ -2364,6 +2400,13 @@ export function SettingsView({
             onRestart={restartViaSettingsSurface}
             isRestarting={isRestarting || hostEngineApplying}
             requiresRestartPending={pendingRestartSections.runtime}
+            computerControlEnabled={settings.computer_use?.enabled ?? false}
+            computerUseBackend={settings.computer_use?.backend ?? "desktop"}
+            computerUseFeature={featureCatalog.find((feature) => feature.name === "computer-use")}
+            computerUseSaving={computerUseSaving === "computer"}
+            computerUseInstalling={nanobotFeatureAction === "enable:computer-use"}
+            capabilityError={nanobotFeaturesError}
+            onToggleComputerControl={(enabled) => void setComputerUseEnabled("computer", enabled)}
           />
         );
       default:
@@ -5255,6 +5298,11 @@ function WebSettings({
   olostepFeature,
   olostepInstalling,
   capabilityError,
+  browserAutomationEnabled,
+  computerUseFeature,
+  computerUseSaving,
+  computerUseInstalling,
+  onToggleBrowserAutomation,
 }: {
   settings: SettingsPayload;
   form: WebSearchSettingsUpdate;
@@ -5274,6 +5322,11 @@ function WebSettings({
   olostepFeature?: NanobotFeatureInfo;
   olostepInstalling: boolean;
   capabilityError: string | null;
+  browserAutomationEnabled: boolean;
+  computerUseFeature?: NanobotFeatureInfo;
+  computerUseSaving: boolean;
+  computerUseInstalling: boolean;
+  onToggleBrowserAutomation: (enabled: boolean) => void;
 }) {
   const { t } = useTranslation();
   const tx = (key: string, fallback: string) => t(key, { defaultValue: fallback });
@@ -5302,9 +5355,46 @@ function WebSettings({
       : selectedProvider?.credential === "base_url"
         ? !baseUrl
         : false;
+  const computerUseInstalled = computerUseFeature?.installed ?? true;
+  const browserAutomationDescription = computerUseInstalling
+    ? tx("settings.help.computerUseInstalling", "Installing computer-use support...")
+    : computerUseInstalled
+      ? tx(
+          "settings.help.browserAutomation",
+          "Let nanobot navigate and act on web pages using structured page elements. Requires Playwright Chromium.",
+        )
+      : tx(
+          "settings.help.computerUseInstall",
+          "Required Python support will be installed when you turn this on.",
+        );
 
   return (
     <div className="space-y-7">
+      <section>
+        <SettingsSectionTitle>
+          {tx("settings.sections.browserAutomation", "Browser automation")}
+        </SettingsSectionTitle>
+        <SettingsGroup>
+          <SettingsRow
+            title={tx("settings.rows.browserAutomation", "Browser automation")}
+            description={browserAutomationDescription}
+          >
+            <ToggleButton
+              checked={browserAutomationEnabled}
+              disabled={computerUseSaving || computerUseInstalling}
+              onChange={onToggleBrowserAutomation}
+              ariaLabel={tx("settings.rows.browserAutomation", "Browser automation")}
+              label={browserAutomationEnabled
+                ? tx("settings.values.on", "On")
+                : tx("settings.values.off", "Off")}
+            />
+          </SettingsRow>
+        </SettingsGroup>
+        {capabilityError ? (
+          <p className="mt-2 px-1 text-[12px] text-destructive">{capabilityError}</p>
+        ) : null}
+      </section>
+
       <section>
         <SettingsSectionTitle>{tx("settings.sections.webSearch", "Web search")}</SettingsSectionTitle>
         {form.provider === "olostep" && olostepFeature && !olostepFeature.installed ? (
@@ -8765,6 +8855,13 @@ function AdvancedSettings({
   onSave,
   onRestart,
   isRestarting,
+  computerControlEnabled,
+  computerUseBackend,
+  computerUseFeature,
+  computerUseSaving,
+  computerUseInstalling,
+  capabilityError,
+  onToggleComputerControl,
 }: {
   form: NetworkSafetySettingsUpdate;
   dirty: boolean;
@@ -8775,11 +8872,60 @@ function AdvancedSettings({
   onSave: () => void;
   onRestart?: () => void;
   isRestarting?: boolean;
+  computerControlEnabled: boolean;
+  computerUseBackend: "desktop" | "browser";
+  computerUseFeature?: NanobotFeatureInfo;
+  computerUseSaving: boolean;
+  computerUseInstalling: boolean;
+  capabilityError: string | null;
+  onToggleComputerControl: (enabled: boolean) => void;
 }) {
   const { t } = useTranslation();
   const tx = (key: string, fallback: string) => t(key, { defaultValue: fallback });
+  const computerUseInstalled = computerUseFeature?.installed ?? true;
+  const computerControlDescription = computerUseInstalling
+    ? tx("settings.help.computerUseInstalling", "Installing computer-use support...")
+    : !computerUseInstalled
+      ? tx(
+          "settings.help.computerUseInstall",
+          "Required Python support will be installed when you turn this on.",
+        )
+      : computerUseBackend === "browser"
+        ? tx(
+            "settings.help.computerControlBrowser",
+            "Pixel-based control currently targets an isolated browser, as configured in config.json.",
+          )
+        : tx(
+            "settings.help.computerControl",
+            "Let nanobot see and control the computer running its engine. macOS requires Screen Recording and Accessibility access.",
+          );
   return (
     <div className="space-y-7">
+      <section>
+        <SettingsSectionTitle>
+          {tx("settings.sections.computerControl", "Computer control")}
+        </SettingsSectionTitle>
+        <SettingsGroup>
+          <SettingsRow
+            title={tx("settings.rows.computerControl", "Computer control")}
+            description={computerControlDescription}
+          >
+            <ToggleButton
+              checked={computerControlEnabled}
+              disabled={computerUseSaving || computerUseInstalling}
+              onChange={onToggleComputerControl}
+              ariaLabel={tx("settings.rows.computerControl", "Computer control")}
+              label={computerControlEnabled
+                ? tx("settings.values.on", "On")
+                : tx("settings.values.off", "Off")}
+            />
+          </SettingsRow>
+        </SettingsGroup>
+        {capabilityError ? (
+          <p className="mt-2 px-1 text-[12px] text-destructive">{capabilityError}</p>
+        ) : null}
+      </section>
+
       <section>
         <SettingsSectionTitle>
           {isNativeHostSurface

webui/src/i18n/locales/en/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "Defaults",
       "webSearch": "Web search",
       "webBehavior": "Behavior",
+      "browserAutomation": "Browser automation",
+      "computerControl": "Computer control",
       "cliApps": "CLI apps",
       "mcp": "MCP services",
       "regional": "Regional",
@@ -199,6 +201,8 @@
       "maxResults": "Max results",
       "timeout": "Timeout",
       "jinaReader": "Jina reader",
+      "browserAutomation": "Browser automation",
+      "computerControl": "Computer control",
       "imageGeneration": "Image generation",
       "imageProvider": "Image provider",
       "imageProviderStatus": "Provider status",
@@ -244,6 +248,11 @@
       "maxResults": "Results returned by each web_search call.",
       "timeout": "Seconds before a search provider request times out.",
       "jinaReader": "Use Jina Reader for web_fetch when available.",
+      "browserAutomation": "Let nanobot navigate and act on web pages using structured page elements. Requires Playwright Chromium.",
+      "computerControl": "Let nanobot see and control the computer running its engine. macOS requires Screen Recording and Accessibility access.",
+      "computerControlBrowser": "Pixel-based control currently targets an isolated browser, as configured in config.json.",
+      "computerUseInstall": "Required Python support will be installed when you turn this on.",
+      "computerUseInstalling": "Installing computer-use support...",
       "imageGeneration": "Expose generate_image in chats when a configured image provider is available.",
       "imageProvider": "Choose the registry provider used by generate_image.",
       "imageProviderStatus": "Image generation reuses provider credentials from Providers.",

webui/src/i18n/locales/es/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "Valores predeterminados",
       "webSearch": "Búsqueda web",
       "webBehavior": "Comportamiento",
+      "browserAutomation": "Automatización del navegador",
+      "computerControl": "Control del ordenador",
       "regional": "Configuración regional",
       "webuiSafety": "Seguridad de WebUI",
       "capabilities": "Capacidades",
@@ -145,6 +147,8 @@
       "maxResults": "Resultados máximos",
       "timeout": "Tiempo de espera",
       "jinaReader": "Lector Jina",
+      "browserAutomation": "Automatización del navegador",
+      "computerControl": "Control del ordenador",
       "imageGeneration": "Generación de imágenes",
       "imageProvider": "Proveedor de imágenes",
       "imageProviderStatus": "Estado del proveedor",
@@ -188,6 +192,11 @@
       "maxResults": "Resultados devueltos por cada llamada web_search.",
       "timeout": "Segundos antes de que una solicitud de búsqueda expire.",
       "jinaReader": "Usa Jina Reader para web_fetch cuando esté disponible.",
+      "browserAutomation": "Permite que nanobot navegue y actúe en páginas web mediante elementos estructurados. Requiere Chromium de Playwright.",
+      "computerControl": "Permite que nanobot vea y controle el ordenador que ejecuta su motor. macOS requiere permisos de grabación de pantalla y accesibilidad.",
+      "computerControlBrowser": "El control por píxeles apunta actualmente a un navegador aislado, según config.json.",
+      "computerUseInstall": "El soporte de Python necesario se instalará al activarlo.",
+      "computerUseInstalling": "Instalando componentes de control...",
       "imageGeneration": "Expone generate_image en chats cuando hay un proveedor de imagen configurado.",
       "imageProvider": "Elige el proveedor registrado usado por generate_image.",
       "imageProviderStatus": "La generación de imágenes reutiliza las credenciales de los proveedores.",

webui/src/i18n/locales/fr/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "Valeurs par défaut",
       "webSearch": "Recherche web",
       "webBehavior": "Comportement",
+      "browserAutomation": "Automatisation du navigateur",
+      "computerControl": "Contrôle de l’ordinateur",
       "regional": "Paramètres régionaux",
       "webuiSafety": "Sécurité WebUI",
       "capabilities": "Capacités",
@@ -145,6 +147,8 @@
       "maxResults": "Résultats max.",
       "timeout": "Délai d’attente",
       "jinaReader": "Lecteur Jina",
+      "browserAutomation": "Automatisation du navigateur",
+      "computerControl": "Contrôle de l’ordinateur",
       "imageGeneration": "Génération d’images",
       "imageProvider": "Fournisseur d’images",
       "imageProviderStatus": "État du fournisseur",
@@ -188,6 +192,11 @@
       "maxResults": "Résultats renvoyés par chaque appel web_search.",
       "timeout": "Nombre de secondes avant l’expiration d’une requête de recherche.",
       "jinaReader": "Utilise Jina Reader pour web_fetch lorsque disponible.",
+      "browserAutomation": "Permet à nanobot de parcourir et manipuler les pages web à partir de leurs éléments structurés. Nécessite Chromium de Playwright.",
+      "computerControl": "Permet à nanobot de voir et contrôler l’ordinateur qui exécute son moteur. macOS exige les autorisations Enregistrement de l’écran et Accessibilité.",
+      "computerControlBrowser": "Le contrôle par pixels cible actuellement un navigateur isolé, conformément à config.json.",
+      "computerUseInstall": "Les composants Python requis seront installés lors de l’activation.",
+      "computerUseInstalling": "Installation des composants de contrôle...",
       "imageGeneration": "Expose generate_image dans les chats lorsqu’un fournisseur d’image configuré est disponible.",
       "imageProvider": "Choisissez le fournisseur inscrit utilisé par generate_image.",
       "imageProviderStatus": "La génération d’images réutilise les identifiants des fournisseurs.",

webui/src/i18n/locales/id/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "Bawaan",
       "webSearch": "Pencarian web",
       "webBehavior": "Perilaku",
+      "browserAutomation": "Otomatisasi browser",
+      "computerControl": "Kontrol komputer",
       "regional": "Regional",
       "webuiSafety": "Keamanan WebUI",
       "capabilities": "Kemampuan",
@@ -145,6 +147,8 @@
       "maxResults": "Hasil maksimum",
       "timeout": "Batas waktu",
       "jinaReader": "Pembaca Jina",
+      "browserAutomation": "Otomatisasi browser",
+      "computerControl": "Kontrol komputer",
       "imageGeneration": "Pembuatan gambar",
       "imageProvider": "Penyedia gambar",
       "imageProviderStatus": "Status penyedia",
@@ -188,6 +192,11 @@
       "maxResults": "Hasil yang dikembalikan oleh setiap panggilan web_search.",
       "timeout": "Detik sebelum permintaan penyedia pencarian mencapai batas waktu.",
       "jinaReader": "Gunakan Jina Reader untuk web_fetch jika tersedia.",
+      "browserAutomation": "Izinkan nanobot menjelajah dan bertindak pada halaman web melalui elemen terstruktur. Memerlukan Chromium dari Playwright.",
+      "computerControl": "Izinkan nanobot melihat dan mengontrol komputer yang menjalankan mesinnya. macOS memerlukan izin Perekaman Layar dan Aksesibilitas.",
+      "computerControlBrowser": "Kontrol berbasis piksel saat ini menargetkan browser terisolasi sesuai config.json.",
+      "computerUseInstall": "Dukungan Python yang diperlukan akan dipasang saat diaktifkan.",
+      "computerUseInstalling": "Memasang komponen kontrol komputer...",
       "imageGeneration": "Tampilkan generate_image di chat saat penyedia gambar yang dikonfigurasi tersedia.",
       "imageProvider": "Pilih penyedia registry yang digunakan oleh generate_image.",
       "imageProviderStatus": "Pembuatan gambar menggunakan kembali kredensial penyedia dari bagian Penyedia.",

webui/src/i18n/locales/ja/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "既定値",
       "webSearch": "ウェブ検索",
       "webBehavior": "動作",
+      "browserAutomation": "ブラウザ自動操作",
+      "computerControl": "コンピュータ操作",
       "regional": "地域",
       "webuiSafety": "WebUI の安全性",
       "capabilities": "機能",
@@ -145,6 +147,8 @@
       "maxResults": "最大結果数",
       "timeout": "タイムアウト",
       "jinaReader": "Jina リーダー",
+      "browserAutomation": "ブラウザ自動操作",
+      "computerControl": "コンピュータ操作",
       "imageGeneration": "画像生成",
       "imageProvider": "画像プロバイダー",
       "imageProviderStatus": "プロバイダー状態",
@@ -188,6 +192,11 @@
       "maxResults": "各 web_search 呼び出しで返す結果数です。",
       "timeout": "検索プロバイダーのリクエストがタイムアウトするまでの秒数です。",
       "jinaReader": "利用可能な場合、web_fetch に Jina Reader を使います。",
+      "browserAutomation": "構造化されたページ要素を使って、nanobot がウェブページを閲覧・操作できるようにします。Playwright Chromium が必要です。",
+      "computerControl": "nanobot がエンジンを実行しているコンピュータを表示・操作できるようにします。macOS では画面収録とアクセシビリティの許可が必要です。",
+      "computerControlBrowser": "ピクセル操作は現在、config.json の設定に従って分離ブラウザを対象としています。",
+      "computerUseInstall": "有効にすると必要な Python コンポーネントがインストールされます。",
+      "computerUseInstalling": "コンピュータ操作コンポーネントをインストール中...",
       "imageGeneration": "画像プロバイダーが設定済みのとき、チャットで generate_image を有効にします。",
       "imageProvider": "generate_image で使用する登録済みプロバイダーを選択します。",
       "imageProviderStatus": "画像生成はプロバイダー設定の認証情報を再利用します。",

webui/src/i18n/locales/ko/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "기본값",
       "webSearch": "웹 검색",
       "webBehavior": "동작",
+      "browserAutomation": "브라우저 자동화",
+      "computerControl": "컴퓨터 제어",
       "regional": "지역",
       "webuiSafety": "WebUI 보안",
       "capabilities": "기능",
@@ -145,6 +147,8 @@
       "maxResults": "최대 결과 수",
       "timeout": "타임아웃",
       "jinaReader": "Jina 리더",
+      "browserAutomation": "브라우저 자동화",
+      "computerControl": "컴퓨터 제어",
       "imageGeneration": "이미지 생성",
       "imageProvider": "이미지 제공자",
       "imageProviderStatus": "제공자 상태",
@@ -188,6 +192,11 @@
       "maxResults": "각 web_search 호출에서 반환되는 결과 수입니다.",
       "timeout": "검색 제공자 요청이 타임아웃되기 전의 초입니다.",
       "jinaReader": "가능할 때 web_fetch에 Jina Reader를 사용합니다.",
+      "browserAutomation": "nanobot이 구조화된 페이지 요소를 사용해 웹페이지를 탐색하고 조작할 수 있게 합니다. Playwright Chromium이 필요합니다.",
+      "computerControl": "nanobot이 엔진을 실행하는 컴퓨터를 보고 제어할 수 있게 합니다. macOS에서는 화면 기록 및 손쉬운 사용 권한이 필요합니다.",
+      "computerControlBrowser": "픽셀 기반 제어는 현재 config.json 설정에 따라 격리된 브라우저를 대상으로 합니다.",
+      "computerUseInstall": "켜면 필요한 Python 구성 요소가 설치됩니다.",
+      "computerUseInstalling": "컴퓨터 제어 구성 요소 설치 중...",
       "imageGeneration": "구성된 이미지 제공자가 있을 때 채팅에서 generate_image를 노출합니다.",
       "imageProvider": "generate_image에 사용할 등록 제공자를 선택합니다.",
       "imageProviderStatus": "이미지 생성은 제공자 자격 증명을 재사용합니다.",

webui/src/i18n/locales/pt-BR/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "Padrões",
       "webSearch": "Busca na web",
       "webBehavior": "Comportamento",
+      "browserAutomation": "Automação do navegador",
+      "computerControl": "Controle do computador",
       "cliApps": "Aplicativos CLI",
       "mcp": "Servidores MCP",
       "regional": "Regional",
@@ -199,6 +201,8 @@
       "maxResults": "Máx. de resultados",
       "timeout": "Tempo limite",
       "jinaReader": "Leitor Jina",
+      "browserAutomation": "Automação do navegador",
+      "computerControl": "Controle do computador",
       "imageGeneration": "Geração de imagens",
       "imageProvider": "Provedor de imagem",
       "imageProviderStatus": "Status do provedor",
@@ -244,6 +248,11 @@
       "maxResults": "Resultados retornados por cada chamada de web_search.",
       "timeout": "Segundos antes de uma requisição de busca expirar.",
       "jinaReader": "Usa o Jina Reader para web_fetch quando disponível.",
+      "browserAutomation": "Permite que o nanobot navegue e interaja com páginas usando elementos estruturados. Requer o Chromium do Playwright.",
+      "computerControl": "Permite que o nanobot veja e controle o computador que executa o mecanismo. No macOS, requer acesso à Gravação de Tela e Acessibilidade.",
+      "computerControlBrowser": "O controle por pixels está direcionado a um navegador isolado, conforme o config.json.",
+      "computerUseInstall": "O suporte Python necessário será instalado ao ativar.",
+      "computerUseInstalling": "Instalando componentes de controle...",
       "imageGeneration": "Expõe generate_image nas conversas quando há um provedor de imagem configurado.",
       "imageProvider": "Escolha o provedor do registro usado por generate_image.",
       "imageProviderStatus": "A geração de imagens reaproveita as credenciais dos provedores.",

webui/src/i18n/locales/vi/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "Mặc định",
       "webSearch": "Tìm kiếm web",
       "webBehavior": "Hành vi",
+      "browserAutomation": "Tự động hóa trình duyệt",
+      "computerControl": "Điều khiển máy tính",
       "regional": "Khu vực",
       "webuiSafety": "An toàn WebUI",
       "capabilities": "Khả năng",
@@ -145,6 +147,8 @@
       "maxResults": "Kết quả tối đa",
       "timeout": "Thời gian chờ",
       "jinaReader": "Trình đọc Jina",
+      "browserAutomation": "Tự động hóa trình duyệt",
+      "computerControl": "Điều khiển máy tính",
       "imageGeneration": "Tạo hình ảnh",
       "imageProvider": "Nhà cung cấp hình ảnh",
       "imageProviderStatus": "Trạng thái nhà cung cấp",
@@ -188,6 +192,11 @@
       "maxResults": "Số kết quả được trả về sau mỗi lần gọi web_search.",
       "timeout": "Số giây trước khi yêu cầu của nhà cung cấp tìm kiếm hết thời gian.",
       "jinaReader": "Dùng Jina Reader cho web_fetch khi có thể.",
+      "browserAutomation": "Cho phép nanobot duyệt và thao tác trên trang web bằng các phần tử có cấu trúc. Cần Playwright Chromium.",
+      "computerControl": "Cho phép nanobot xem và điều khiển máy tính đang chạy engine. macOS yêu cầu quyền Ghi màn hình và Trợ năng.",
+      "computerControlBrowser": "Điều khiển theo điểm ảnh hiện nhắm tới trình duyệt cách ly theo config.json.",
+      "computerUseInstall": "Hỗ trợ Python cần thiết sẽ được cài đặt khi bật.",
+      "computerUseInstalling": "Đang cài đặt thành phần điều khiển...",
       "imageGeneration": "Hiển thị generate_image trong chat khi đã cấu hình nhà cung cấp hình ảnh.",
       "imageProvider": "Chọn nhà cung cấp registry được generate_image sử dụng.",
       "imageProviderStatus": "Tạo ảnh dùng lại thông tin xác thực từ mục Nhà cung cấp.",

webui/src/i18n/locales/zh-CN/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "默认值",
       "webSearch": "网络搜索",
       "webBehavior": "网络行为",
+      "browserAutomation": "浏览器自动化",
+      "computerControl": "电脑控制",
       "cliApps": "CLI 应用",
       "mcp": "MCP 服务",
       "regional": "区域",
@@ -199,6 +201,8 @@
       "maxResults": "最大结果数",
       "timeout": "超时",
       "jinaReader": "Jina 阅读器",
+      "browserAutomation": "浏览器自动化",
+      "computerControl": "电脑控制",
       "imageGeneration": "图片生成",
       "imageProvider": "图片提供商",
       "imageProviderStatus": "提供商状态",
@@ -244,6 +248,11 @@
       "maxResults": "每次 web_search 调用返回的结果数。",
       "timeout": "搜索提供商请求超时前等待的秒数。",
       "jinaReader": "可用时为 web_fetch 使用 Jina Reader。",
+      "browserAutomation": "允许 nanobot 通过结构化页面元素浏览网页并执行操作，需要 Playwright Chromium。",
+      "computerControl": "允许 nanobot 查看并控制运行引擎的电脑。macOS 需要授予“屏幕录制”和“辅助功能”权限。",
+      "computerControlBrowser": "像素控制当前按 config.json 配置作用于隔离浏览器。",
+      "computerUseInstall": "开启时会自动安装所需 Python 组件。",
+      "computerUseInstalling": "正在安装电脑操作组件...",
       "imageGeneration": "配置图片提供商后，即可在聊天中使用 generate_image。",
       "imageProvider": "选择 generate_image 使用的注册提供商。",
       "imageProviderStatus": "图片生成会复用「提供商」里的凭据。",

webui/src/i18n/locales/zh-TW/common.json

@@ -115,6 +115,8 @@
       "imageDefaults": "預設值",
       "webSearch": "網路搜尋",
       "webBehavior": "網路行為",
+      "browserAutomation": "瀏覽器自動化",
+      "computerControl": "電腦控制",
       "regional": "區域",
       "webuiSafety": "WebUI 安全",
       "capabilities": "能力",
@@ -145,6 +147,8 @@
       "maxResults": "最大結果數",
       "timeout": "逾時",
       "jinaReader": "Jina 閱讀器",
+      "browserAutomation": "瀏覽器自動化",
+      "computerControl": "電腦控制",
       "imageGeneration": "圖片生成",
       "imageProvider": "圖片供應商",
       "imageProviderStatus": "供應商狀態",
@@ -188,6 +192,11 @@
       "maxResults": "每次呼叫 web_search 所回傳的結果數。",
       "timeout": "搜尋供應商請求逾時前的秒數。",
       "jinaReader": "若可用，則讓 web_fetch 使用 Jina Reader。",
+      "browserAutomation": "允許 nanobot 透過結構化頁面元素瀏覽網頁並執行操作，需要 Playwright Chromium。",
+      "computerControl": "允許 nanobot 查看並控制執行引擎的電腦。macOS 需要授予螢幕錄製與輔助使用權限。",
+      "computerControlBrowser": "像素控制目前依 config.json 設定作用於隔離瀏覽器。",
+      "computerUseInstall": "開啟時會自動安裝所需 Python 元件。",
+      "computerUseInstalling": "正在安裝電腦操作元件...",
       "imageGeneration": "設定圖片供應商後，即可在聊天中使用 generate_image。",
       "imageProvider": "選擇 generate_image 使用的註冊供應商。",
       "imageProviderStatus": "圖片生成功能會沿用 [供應商] 中的憑證。",

webui/src/lib/api.ts

@@ -7,6 +7,7 @@ import type {
   ChannelValidationPayload,
   ChatSummary,
   CliAppsPayload,
+  ComputerUseSettingsUpdate,
   FilePreviewPayload,
   ImageGenerationSettingsUpdate,
   McpPresetsPayload,
@@ -1077,6 +1078,24 @@ export async function updateNetworkSafetySettings(
   );
 }
 
+export async function updateComputerUseSettings(
+  token: string,
+  update: ComputerUseSettingsUpdate,
+  base: string = "",
+): Promise<SettingsPayload> {
+  const query = new URLSearchParams();
+  if (update.browserEnabled !== undefined) {
+    query.set("browser_enabled", String(update.browserEnabled));
+  }
+  if (update.computerEnabled !== undefined) {
+    query.set("enabled", String(update.computerEnabled));
+  }
+  return request<SettingsPayload>(
+    `${base}/api/settings/computer-use/update?${query}`,
+    token,
+  );
+}
+
 export async function updateImageGenerationSettings(
   token: string,
   update: ImageGenerationSettingsUpdate,

webui/src/lib/types.ts

@@ -579,6 +579,11 @@ export interface SettingsPayload {
       use_jina_reader: boolean;
     };
   };
+  computer_use?: {
+    browser_enabled: boolean;
+    enabled: boolean;
+    backend: "desktop" | "browser";
+  };
   api?: {
     host: string;
     port: number;
@@ -1102,6 +1107,11 @@ export interface NetworkSafetySettingsUpdate {
   webuiDefaultAccessMode: WebuiDefaultAccessMode;
 }
 
+export interface ComputerUseSettingsUpdate {
+  browserEnabled?: boolean;
+  computerEnabled?: boolean;
+}
+
 export interface ImageGenerationSettingsUpdate {
   enabled: boolean;
   provider: string;

webui/src/tests/api.test.ts

@@ -46,6 +46,7 @@ import {
   pollChannelConnect,
   startChannelConnect,
   updateAutomation,
+  updateComputerUseSettings,
   updateSidebarState,
   updateImageGenerationSettings,
   updateModelCallOrder,
@@ -836,6 +837,20 @@ describe("webui API helpers", () => {
     );
   });
 
+  it("updates computer-use capability switches", async () => {
+    await updateComputerUseSettings("tok", { browserEnabled: true });
+    expect(fetch).toHaveBeenCalledWith(
+      "/api/settings/computer-use/update?browser_enabled=true",
+      expect.objectContaining({ headers: { Authorization: "Bearer tok" } }),
+    );
+
+    await updateComputerUseSettings("tok", { computerEnabled: false });
+    expect(fetch).toHaveBeenCalledWith(
+      "/api/settings/computer-use/update?enabled=false",
+      expect.objectContaining({ headers: { Authorization: "Bearer tok" } }),
+    );
+  });
+
   it("manages the API service capability", async () => {
     await fetchApiService("tok");
     expect(fetch).toHaveBeenCalledWith(

webui/src/tests/i18n.test.tsx

@@ -69,6 +69,8 @@ const LOCALIZED_SETTINGS_COPY_KEYS = [
   "settings.sections.localPreferences",
   "settings.sections.webSearch",
   "settings.sections.webBehavior",
+  "settings.sections.browserAutomation",
+  "settings.sections.computerControl",
   "settings.sections.webuiSafety",
   "settings.sections.capabilities",
   "settings.sections.apps",
@@ -149,6 +151,8 @@ const LOCALIZED_SETTINGS_COPY_KEYS = [
   "settings.rows.fileEditDisplay",
   "settings.rows.codeWrap",
   "settings.rows.brandLogos",
+  "settings.rows.browserAutomation",
+  "settings.rows.computerControl",
   "settings.rows.currentModel",
   "settings.rows.localServiceAccess",
   "settings.rows.webuiDefaultAccess",
@@ -160,6 +164,11 @@ const LOCALIZED_SETTINGS_COPY_KEYS = [
   "settings.help.fileEditDisplay",
   "settings.help.codeWrap",
   "settings.help.brandLogos",
+  "settings.help.browserAutomation",
+  "settings.help.computerControl",
+  "settings.help.computerControlBrowser",
+  "settings.help.computerUseInstall",
+  "settings.help.computerUseInstalling",
   "settings.help.currentModel",
   "settings.help.localServiceAccess",
   "settings.help.webuiDefaultAccess",

webui/src/tests/settings-view.test.tsx

@@ -64,6 +64,11 @@ function settingsPayload(): SettingsPayload {
       search: { max_results: 5, timeout: 30 },
       fetch: { use_jina_reader: true },
     },
+    computer_use: {
+      browser_enabled: false,
+      enabled: false,
+      backend: "desktop",
+    },
     api: {
       host: "127.0.0.1",
       port: 8900,
@@ -4202,6 +4207,98 @@ describe("SettingsView Apps catalog", () => {
     });
   });
 
+  it("enables browser automation from Web settings", async () => {
+    const payload = settingsPayload();
+    const updatedPayload: SettingsPayload = {
+      ...payload,
+      computer_use: { ...payload.computer_use!, browser_enabled: true },
+      requires_restart: true,
+      restart_required_sections: ["runtime"],
+    };
+    const fetchMock = vi.fn(async (input: RequestInfo | URL) => {
+      const url = String(input);
+      if (url === "/api/settings") return jsonResponse(payload);
+      if (url === "/api/settings/nanobot-features") {
+        return jsonResponse({
+          features: [{
+            name: "computer-use",
+            display_name: "Computer Use",
+            type: "feature",
+            enabled: true,
+            installed: true,
+            ready: true,
+            status: "enabled",
+            install_supported: true,
+            requires_restart: true,
+          }],
+          enabled_count: 1,
+        });
+      }
+      if (url === "/api/settings/computer-use/update?browser_enabled=true") {
+        return jsonResponse(updatedPayload);
+      }
+      return { ok: false, status: 404, json: async () => ({}) } as Response;
+    });
+    vi.stubGlobal("fetch", fetchMock);
+
+    renderSettingsView({ initialSection: "browser" });
+
+    fireEvent.click(await screen.findByRole("switch", { name: "Browser automation" }));
+
+    await waitFor(() =>
+      expect(fetchMock).toHaveBeenCalledWith(
+        "/api/settings/computer-use/update?browser_enabled=true",
+        expect.objectContaining({ headers: { Authorization: "Bearer tok" } }),
+      ),
+    );
+  });
+
+  it("enables computer control from Security settings", async () => {
+    const payload = settingsPayload();
+    const updatedPayload: SettingsPayload = {
+      ...payload,
+      computer_use: { ...payload.computer_use!, enabled: true },
+      requires_restart: true,
+      restart_required_sections: ["runtime"],
+    };
+    const fetchMock = vi.fn(async (input: RequestInfo | URL) => {
+      const url = String(input);
+      if (url === "/api/settings") return jsonResponse(payload);
+      if (url === "/api/settings/nanobot-features") {
+        return jsonResponse({
+          features: [{
+            name: "computer-use",
+            display_name: "Computer Use",
+            type: "feature",
+            enabled: true,
+            installed: true,
+            ready: true,
+            status: "enabled",
+            install_supported: true,
+            requires_restart: true,
+          }],
+          enabled_count: 1,
+        });
+      }
+      if (url === "/api/settings/computer-use/update?enabled=true") {
+        return jsonResponse(updatedPayload);
+      }
+      return { ok: false, status: 404, json: async () => ({}) } as Response;
+    });
+    vi.stubGlobal("fetch", fetchMock);
+
+    renderSettingsView({ initialSection: "advanced" });
+
+    fireEvent.click(await screen.findByRole("switch", { name: "Computer control" }));
+
+    await waitFor(() =>
+      expect(fetchMock).toHaveBeenCalledWith(
+        "/api/settings/computer-use/update?enabled=true",
+        expect.objectContaining({ headers: { Authorization: "Bearer tok" } }),
+      ),
+    );
+  });
+
   it("saves network safety without exposing technical SSRF copy", async () => {
     const payload = settingsPayload();
     const fetchMock = vi.fn(async (input: RequestInfo | URL) => {