AmrDab/clawdcursor

npm i -g clawdcursor
clawdcursor consent --accept    # one-time desktop-control consent (required)
clawdcursor grant                # macOS only — approve Accessibility + Screen Recording
claude mcp add clawdcursor -s user -- clawdcursor mcp --compact   # Claude Code
claude plugin marketplace add AmrDab/clawdcursor && claude plugin install clawdcursor@clawdcursor   # one-command plugin

Usage

• •Add to any MCP-capable agent: Claude Code (`clawdcursor mcp --compact`), Codex (`args = ["mcp", "--compact"]` in ~/.codex/config.toml), Cursor/Windsurf/Claude Desktop (mcpServers), or Zed (`context_servers`).
• •Ask the agent to operate apps — e.g. "open Outlook and reply to the latest email from Sarah"; the agent acts on elements by {element_id, snapshot_id} instead of pixels.
• •Optional `expect` on a consequential action makes the server re-check the live screen and report a DEVIATION instead of a hollow success.
• •Headless/HTTP mode: run `clawdcursor agent` for a daemon exposing HTTP MCP at http://127.0.0.1:3847/mcp with a bearer token at ~/.clawdcursor/token.

Key features

• ✓Cheapest-tier-first perception: a11y tree (free) → OCR (cheap) → DOM (CDP) → screenshot/vision (expensive, last resort), so token cost tracks task difficulty.
• ✓Reactive verification via `expect` — actions are checked against fresh observations on the live screen.
• ✓Single safety gate (`safety.evaluate()` allow/confirm/block) before anything touches the desktop, plus an on-screen 'desktop control in progress' banner with a double-click-to-stop kill switch.
• ✓98 granular tools + 7 compound tools (computer/accessibility/window/system/browser/task/batch) through one MCP config; batch collapses N steps into one call.
• ✓Cross-platform PlatformAdapter for Windows, macOS, Linux X11 and Wayland; stdio and HTTP MCP transports; secure-field redaction and no telemetry by default.

Best for

Last-mile GUI automation for AI agents over native desktop apps, legacy apps, and canvas-only UIs that have no public API or CLI — acting on elements by accessibility id rather than pixels.

Caveats

• ⚠macOS requires Accessibility permission (and optional Screen Recording for vision fallback); Xcode Command Line Tools needed for screenshots/vision.
• ⚠Linux needs system packages: tesseract-ocr, python3-gi + gir1.2-atspi-2.0, and on Wayland also ydotool (synthetic input).
• ⚠Requires Node.js 20+; one-time desktop-control consent is mandatory.
• ⚠Blocked sequences (Ctrl+Alt+Del, lock/log-out/force-quit/shutdown) are refused outright; sensitive apps auto-elevate to Confirm.
• ⚠HTTP mode binds to 127.0.0.1 with bearer-token auth; screenshots stay in RAM and only go to the model endpoint you configured.