High-risk tools in CoWork OS

Render a camera view to an image

Execute a batch of browser actions in sequence. Use for multi-step interactions (e.g. fill form, click submit, wait for result).

Attach to an existing Chrome browser session via Chrome DevTools Protocol.

Go back in browser history

Click on an element on the page

Close the browser

Close a browser tab by tab id

Drag from one Browser V2 snapshot ref to another

Set the visible in-app browser viewport/device emulation for responsive testing.

Execute JavaScript code in the browser context

Fill a form field with text

Go forward in browser history

Accept or dismiss the latest JavaScript dialog in the browser

Move the browser pointer over an element, preferably by browser_snapshot ref

Navigate the browser to a URL. By default, CoWork uses background/headless browser control unless settings or the tool input request a visible workbench.

Press a keyboard key (e.g., Enter, Tab, Escape)

Reload the current page

Scroll the page

Select an option from a dropdown

Switch to a browser tab by tab id

Start a lightweight browser trace for diagnostics

Stop the active lightweight browser trace

Type text character by character (useful for search boxes with autocomplete)

Wait for an element to appear on the page

Execute JavaScript code in the canvas context.

Open a remote web page inside the canvas window for full in-app browsing.

OPTIONAL: Open the canvas in a separate interactive window.

Click inside the current controlled window using screenshot-relative coordinates.

Create a new cloud sandbox (Linux VM). Returns sandbox ID and connection info.

Execute a shell command in a cloud sandbox. Returns stdout, stderr, and exit code.

Submit a Flux-style photoreal workflow graph with project metadata

Submit a ComfyUI workflow JSON graph

Compile a workspace .tex file into a PDF using a system LaTeX engine.

Double-click inside the current controlled window using screenshot-relative coordinates. Returns a fresh screenshot.

Drag through a series of screenshot-relative points inside the current controlled window. Returns a fresh screenshot.

Execute Python, JavaScript, or shell code in a sandboxed process with a configurable timeout.

Generate an image from a text description using AI. CoWork OS will pick the best configured provider by default (Gemini/OpenAI/ChatGPT Subscription/Azure/OpenRouter), unless you...

Generate narrated MP3 audio from text using the configured voice service.

Generate a video from a text prompt (or an image/video reference) using AI. CoWork OS will pick the best configured provider automatically (OpenAI Sora 2 / Azure Sora 2 / Gemini...

Merge the current worktree branch back to the base branch. Only available in worktree isolation mode. Use when your work is complete and ready to be integrated.

Run raw Google Slides presentations.batchUpdate requests for advanced edits

Make HTTP requests like curl. Supports all HTTP methods, custom headers, and request bodies.

Send a key or key chord to the current controlled window. Modifier keys: cmd/command on macOS, win/windows on Windows, ctrl/control, alt/option, shift. Returns a fresh screenshot.

Mention another specialized agent to request help, delegate work, request a review, or share information.

Run deterministic, sandboxed Python-subset code (Monty) for fast local computation and post-processing.

Run a named workspace transform from .cowork/transforms/<name>.monty. The input object is available as

Read a text file, run a named workspace transform, and write the result to an output file, returning only a small summary.

Move the pointer inside the current controlled window using screenshot-relative coordinates. Returns a fresh screenshot.

Take a photo using a mobile companion

Send an SMS message using an Android mobile companion device. Not available on iOS.

Open an application by name (e.g.,

Open a file or folder with the system default application

Open a URL in the default web browser

Spawn multiple sub-agents in parallel and wait for all of them to complete.

Pause a running descendant child agent task. This tool only works for tasks spawned by the current task (descendants).

Run a specific QA check on the current page.

Perform an interaction on the current QA page (click, fill, hover, scroll, wait).

Navigate the QA browser to a URL and get a page state report.

Run a full automated QA pipeline on a web app using Playwright.

Resume a paused descendant child agent task. This tool only works for tasks spawned by the current task (descendants).

Generate building massing from the brief and constraints

Execute exact AppleScript / osascript code on macOS.

Execute a shell command in the workspace directory. IMPORTANT: This tool requires user approval before execution. The user will see the command and can approve or deny it. Use t...

Schedule a task to run at a specific time or on a recurring basis.

Run a multi-step scraping session with persistent browser state.

Scroll at a screenshot-relative point inside the current controlled window. scrollX and scrollY are signed line deltas. Returns a fresh screenshot.

Show a file in the system file manager (Finder on macOS, Explorer on Windows)

Invoke a skill by ID when one clearly matches the task. If a matching skill exists, call Skill before using other tools or drafting the final answer.

Spawn a new agent (sub-task) to work on a specific task independently. Use this to delegate work,

Switch to a different workspace/working directory. Use this when you need to work in a different folder

Type or set text into the currently focused control of the current controlled window. Returns a fresh screenshot.

Initiate an outbound phone call via ElevenLabs Agents + Twilio integration.

Pause briefly and then refresh the current controlled window screenshot. Useful after async UI transitions.

Wait for a spawned agent to complete and retrieve its results. Returns the agent