Clawd Cursor: desktop control through MCP
Clawd Cursor
An MCP server that lets an agent drive windows and apps: it reads the screen as structure and acts by element name
- 406 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent gets control of a real desktop: it clicks, types and reads windows, including mail and messengers
- Once granted, the consent and the Accessibility and Screen Recording permissions on macOS open access to every app on the machine
- Launching with npx -y fetches the latest package version every time, so the author advises a global install for a tool like this
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/clawd-cursorDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent gets control of a real desktop: it clicks, types and reads windows, including mail and messengers
- Once granted, the consent and the Accessibility and Screen Recording permissions on macOS open access to every app on the machine
- Launching with npx -y fetches the latest package version every time, so the author advises a global install for a tool like this
This is third-party code. Review the repository files before installing.
What it does
Clawd Cursor runs locally and lets an agent drive a real desktop: open apps, click, type, and read windows and browser pages. It fits where a service has neither an API nor a command line: a legacy app, a window with no integration, an interface where all that is left is pressing a button. The server turns the screen into one interface map by merging the accessibility tree with text recognition, and gives every element an identifier of the form el_NN. The agent acts by identifier, and pixel coordinates are needed only at the last step. A screenshot with a vision model is the last step, for windows without structure such as canvases and games. An action can carry an expect parameter, and then the server checks the screen and returns DEVIATION if the interface did not respond as expected. Every call goes through a single safety checkpoint with a verdict: allow, ask for confirmation or block. While the agent works, the screen shows a banner and an amber glow around the edges.
Who it is for. For people who hand routine actions in graphical programs on Windows, macOS or Linux to an agent when no suitable API exists.
Good fit when
- You need to reply to an email or fill in a form in an app that has no API or command line
- A legacy desktop app has to be plugged into an agent chain without an integration
- You need to make sure the interface really changed after a button press
- You want to avoid a screenshot on every step: window structure and text recognition come first
- The agent must stop at a sign-in or a code prompt and wait for you
Not a fit when
- The service has an API or a command line: the author advises using those and keeping Clawd Cursor as a fallback
- The computer has open banking sessions or other people's data that the agent must not see
- You only need browser automation for tests: Playwright and similar tools fit that
- You are not ready to watch the agent work: it presses buttons in real programs
Example request
Open Outlook and reply to Sarah's latest email that the call moves to Thursday. Show me the text before sending.MCP
- Transport
- stdio, http
- Authentication
- not required
Limitations
Before the first use you must give consent to control the computer once with the consent --accept command. Hosts that can ask the user a question over MCP, such as Claude Code and VS Code, ask for it on the first tool call. Node.js 20 or newer is required.
On macOS the Accessibility permission is mandatory, while Screen Recording is needed only for screenshots and text recognition, and the clawdcursor grant command walks you through it. On Linux you have to install tesseract-ocr, python3-gi and gir1.2-atspi-2.0, and on Wayland ydotool as well. The built-in agent loop with its own model is enabled separately with the clawdcursor agent command and configured through doctor.
How to disable. Remove the server from the client, for example with claude mcp remove clawdcursor, or disable the Claude Code plugin through /plugin. A double click on the on-screen banner stops the agent's work. The global package is removed with npm uninstall -g clawdcursor.
FAQ
How do I install Clawd Cursor in Claude Code?
Run npm i -g clawdcursor, then clawdcursor consent --accept and claude mcp add clawdcursor -s user -- clawdcursor mcp --compact. Instead of the commands you can install the plugin: claude plugin marketplace add AmrDab/clawdcursor and claude plugin install clawdcursor@clawdcursor. The plugin registers the tools and brings a skill with usage rules.
How does the agent know what is on screen if it does not look at screenshots?
It reads the accessibility tree first, which is free. If the tree is empty, the operating system's text recognition kicks in. For Chrome, Electron and WebView it reads the DOM through CDP. A screenshot goes to the model only as the last step, for example for a window with a drawn canvas.
What happens when the agent meets a sign-in or an SMS code?
The agent pauses: the glow around the screen edges turns blue and the banner says what to do. You sign in yourself and press Resume to continue from the same place, or End task to stop it. You can also answer with text through the "Something else" button.
How do I stop an agent that controls my computer?
Double click the banner shown on screen the whole time the agent works. It stops all actions on Windows, macOS and Linux.
Does Clawd Cursor work with only one model?
No, it works with any model that can call tools: Claude, GPT, Gemini or Llama. The server does not call a model itself until you enable the built-in agent loop.
Discussion
No questions or reviews yet.
Related
MCP serversOpen Computer Use
open-computer-use
An open MCP server for desktop control: the agent reads app state and clicks, types and scrolls on macOS, Linux and Windows
An MCP server for driving Windows: the agent opens apps, clicks and types, reads window and UI state, and works with files and PowerShell
An MCP server for macOS: the agent moves the mouse, types, manages windows and captures the screen when an app has no API
An MCP server for Windows: the agent moves the mouse, presses keys, manages windows and takes screenshots
A Swift MCP server that lets an agent drive any macOS app: click, type and read the interface tree
A macOS CLI and MCP server: screenshots, accessibility-based UI inspection and app control via clicks and typing
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Clawd Cursor: desktop control through MCP (this entry) | no | not required | 406 | |
| Windows-MCP | no | not required | 10,740 | controls Windows only, while Clawd Cursor runs on Windows, macOS and Linux |
| MCPControl | no | not required | 332 | moves the mouse, presses keys and takes screenshots on Windows, with no element map or result check |
README in short
The README compares the project with browser-use, Playwright MCP, OmniParser and Anthropic's computer use mode, then gives installation for Claude Desktop (an MCPB extension), Claude Code, Cursor, VS Code, Zed, Codex and opencode. After that come the run modes, a list of seven compound tools, a table of perception tiers and platform support by operating system. A changelog of recent versions sits at the top.