Clawd Cursor

Community

Safe cross-OS desktop control for any AI agent — the fallback execution layer.

How to install ↓ Compare GitHub source repository ↗

Category: Other #20 of 230
Stars: ★ 405 Very popular
Transport: stdio (local process) · Streamable HTTP
Runtime: Node.js 20+
Credentials: No credential needed
License: MIT
Last commit: today
Tools: 14

68FMRS · C

Clawd Cursor is a powerful MCP server that provides safe desktop automation via accessibility tree and OCR, with verification and a single safety gate. It's suitable for agents needing to control native apps, but requires proper permissions and may have a learning curve.

Strongest · Documentation 16/20 Weakest · Reliability 10/20

Reliability

10/20

Security and permissions

13/20

Maintenance

15/20

Documentation

16/20

Setup experience

14/20

Why each score

Reliability 10/20: Evidence shows a TypeScript server with a CI workflow (cross-platform.yml) and claimed 1000+ tests, but lacks specific test details for MCP handshake or key paths. Static review cannot execute, and test evidence is insufficient, so reliability is below 12. Deductions for not showing actual test files and CI logs, and for limited error-handling and mocking evidence.
Security and permissions 13/20: No malicious behavior found, but security relies on descriptions rather than verifiable source. Security design includes binding to 127.0.0.1, bearer token, confirmation for sensitive operations, sensitive app policy, prompt injection defense, and no telemetry by default, but permission and confirmation boundaries are not verified in source. Deductions for security mechanisms being only documented, lack of implementation details, and no security audit evidence. No red lines triggered.
Maintenance 15/20: Repository has 392 stars, MIT license, and security reporting channel (SECURITY.md). But no static evidence of commit frequency, contribution history, or dependency update strategy. npm version and GitHub actions exist but unverified. Deductions for missing maintenance activity data such as commit frequency and response times.
Documentation 16/20: README is detailed and layered, covering installation for Windows/macOS/Linux, transports, tool catalog, safety, platform support, CLI, and development. Includes config examples and troubleshooting hints, but lacks like known limitations or deeper troubleshooting. Deductions for cost estimates and detailed error scenarios not explicitly stated.
Setup experience 14/20: Installation steps are clear: global npm install, consent, grant, and config examples for multiple clients. However, static calibration limits to 15 because no execution verification. Multiple platforms require extra dependencies (e.g., Linux tesseract and ydotool), which are documented. Deductions for inability to verify runtime, and installer scripts provided but not verified.

Static review · not runListed 2026-08-07

Read the FMRS scoring method →

Fit and risk

What it can accessWrites / deletes local filesRuns commands or codeUses the networkControls a browser

Best for

AI agents needing cross-platform desktop automation
Scenarios requiring safety controls with allow/confirm/block
Privacy-sensitive tasks that should stay local
Agents wanting to avoid high vision model costs

Not for

Browser-only automation where specialized tools exist
Cloud-hosted or non-local execution
Tasks requiring strict human approval
Pure pixel-based manipulation without vision models

Required permissions

Desktop control (click, type)
Read screen content (accessibility tree, OCR, screenshots)
Open apps and files
Browser control (CDP)
Clipboard read/write
System-level hotkeys

Risks and side effects

Agent may perform unintended actions, leading to data loss or system changes
Sensitive information on screen could be read by the agent
Potential for malicious actions if agent misbehaves
Requires proper permission grants, otherwise may not work
Confirmation mode may interrupt automation flow

Setup

Before you start

Runtime：Node.js 20+

Other optional settings (1)

CLAWD_LOG optional Set to pretty to show per-call perception-tier badges (obs·a11y/ocr/vision) and act logs live in the terminal; optional.

Install globally: npm i -g clawdcursor
Grant consent: clawdcursor consent --accept
macOS only: grant permissions with clawdcursor grant
Add to MCP config:

{"mcpServers":{"clawdcursor":{"command":"clawdcursor","args":["mcp","--compact"]}}}

claude_desktop_config.json

{
  "mcpServers": {
    "clawdcursor": {
      "command": "clawdcursor",
      "args": [
        "mcp",
        "--compact"
      ]
    }
  }
}

Shown for Claude Desktop. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json

{
  "servers": {
    "clawdcursor": {
      "command": "clawdcursor",
      "args": [
        "mcp",
        "--compact"
      ]
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal

claude mcp add clawdcursor -- clawdcursor mcp --compact

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

Confirm the clawdcursor tools (computer, accessibility, compile_ui, etc.) appear in your client's tool list after adding the config, then ask the agent to run compile_ui and check it returns elements with el_NN ids; one-time setup requires running clawdcursor consent --accept first.

Troubleshooting

Ensure Node.js 20+ is installed
Ensure desktop access permissions have been granted (macOS Accessibility)
Check OS dependencies: Linux requires tesseract-ocr, python3-gi, etc.
Verify the 'clawdcursor' command is in PATH
Check logs: CLAWD_LOG=pretty for detailed operations
If agent can't connect, ensure HTTP mode token is correct

Things to try

Once connected, you can ask your AI assistant things like:

Open Outlook and reply to the latest email from Sarah
Compile the current screen into a UI map and find the Send button, then click it
Open Notepad, type hello, and verify the text was entered
Connect to Chrome and read the text of the current page

Tools 14

computer writes

Performs mouse and keyboard actions like clicking, typing, scrolling, dragging, and waiting.

accessibility read-only

Reads accessibility tree, finds elements, invokes actions, sets values, compiles UI map.

window writes

Manages windows and apps: list, focus, resize, open app, open file/URL, switch tabs.

system read-only

System-level operations: clipboard, OCR, undo, detect webview, run shortcuts.

browser writes

Controls browser via CDP: connect, read text, click, type, evaluate, manage tabs.

task writes

Delegates a task to the built-in thin agent loop.

batch writes

Collapses multiple calls into one round-trip, each step still safety-gated.

compile_ui read-only

Fuses accessibility tree and OCR into a confidence-scored map of UI elements.

Show 6 more tools

find_button read-only

Finds a button by name or text and returns its element ID.

find_field read-only

Finds an input field by name or label and returns its element ID.

smart_click writes

Smart click: tries accessibility tree first, falls back to OCR or vision if needed.

smart_type writes

Smart type: tries accessibility tree first, falls back to OCR or vision if needed.

smart_read read-only

Smart read: tries accessibility tree first, falls back to OCR or vision if needed.

detect_webview read-only

Detects Electron/WebView2 apps and switches to browser control.

Use cases

Driving native apps (e.g., Outlook, Notepad) for complex GUI tasks

Automating browser actions via CDP for Chrome/Edge

Interacting with legacy applications that have no API or CLI

Canvas-based apps when spatial reasoning is required

As a last-resort execution layer after native APIs and CLIs

Supported clients

Claude Desktop

Claude Code

Cursor

Windsurf

Zed

Codex

Listed from the project's documentation, not tested by this site.

Overview

Clawd Cursor is a local MCP server that gives any tool-calling agent (Claude Code, Cursor, Windsurf, OpenClaw, etc.) safe control of the real desktop. It compiles the screen into a UI map by fusing the accessibility tree and OCR, tagging each element with a stable el_NN id, and acts on elements by id rather than pixel coordinates. It verifies its own actions and routes every call through a single safety checkpoint. Works on Windows, macOS, and Linux, fully local with no cloud dependency.

Similar servers

Hevy MCP 77 · B

Manage your Hevy workout data from AI assistants via MCP.

★ 509 · Tools 26 Compare with this →

Firecrawl MCP Server 75 · B

Firecrawl's official MCP server for web search, scraping, and structured extraction for AI agents

★ 7.6k · Tools 11 Compare with this →

BioMCP 75 · B

One binary. One grammar. Evidence from the biomedical sources you already trust.

★ 654 · Tools 11 Compare with this →

Compartment 75 · B

Encrypted, fully offline long-term memory for AI agents

★ 579 · Tools 14 Compare with this →

Compare with other MCP servers

Source revision 295e95c81e34 Data synced 2026-10-12 Read the FMRS scoring method