Clawd Cursor
@AmrDab
The local MCP server that gives any AI agent safe desktop control. Drives Windows, macOS, and Linux GUIs through accessibility, OCR, or vision fallback. 6 compact compound tools covering 97 primitives. Model-agnostic. Local-only by default; bearer-token auth on HTTP; every tool c
概览
What is Clawd Cursor?
Clawd Cursor is a skill—not an app—that gives any tool-calling AI agent the ability to click, type, read the screen, open apps, and drive any GUI on a real desktop, using the Model Context Protocol (MCP). It is model-agnostic, app-agnostic, OS-agnostic, and runs locally by default.
How to use Clawd Cursor?
Install via curl (macOS/Linux) or PowerShell (Windows). Run clawdcursor consent --accept for one-time desktop-control consent, then clawdcursor status to verify. For editor hosts (Claude Code, Cursor, Windsurf, Zed), use clawdcursor mcp as a stdio MCP server. For headless daemons, run clawdcursor agent (HTTP MCP on 127.0.0.1:3847). Wire the command into your host’s MCP configuration.
Key features of Clawd Cursor
- Two MCP transports: stdio for editors, HTTP for daemons.
- Six compound tools (Anthropic
computer_20250124-style) plus granular tools. - Cheapest-tier-first pipeline: accessibility → OCR → screenshot → vision.
- Ground-truth verification with six independent signals and weighted voting.
- Reflector loop that emits structured failure causes and suggested strategies.
- Single safety chokepoint (
safety.evaluate()) on every tool call.
Use cases of Clawd Cursor
- Driving native apps or legacy enterprise tools that have no API.
- Automating web portals behind SSO that block headless browsers.
- Interacting with anything inside a Citrix or RDP session.
- Performing GUI tasks in any application from any tool-calling LLM.
FAQ from Clawd Cursor
What makes Clawd Cursor different from using a native API or CLI?
Clawd Cursor is designed for the “last mile” when no API, CLI, or direct file edit exists. It is slower, less reliable, and more token-hungry than those alternatives, but it works where they cannot—such as legacy ERP, Electron dialogs, or Citrix sessions.
What are the system requirements?
Node.js 20 or later. Windows, macOS, or Linux (X11 and Wayland). Linux additionally requires tesseract-ocr, python3-gi, gir1.2-atspi-2.0, and (Wayland only) ydotool or wtype. macOS must grant Accessibility and Screen Recording permissions via clawdcursor grant.
Where does my data go?
Clawd Cursor runs locally and binds to 127.0.0.1 by default. Screenshots stay in RAM unless you deliberately route them to a cloud model. No telemetry is collected.
What transport options are available?
Two MCP transports: stdio MCP (for editor hosts like Claude Code, Cursor, Windsurf, Zed) and HTTP MCP (for daemons, bring-your-own-agent, or multi-process orchestration). Both are stateless and share the same tool catalog and JSON-RPC envelope.
Can I use Clawd Cursor with any AI model?
Yes. It is model-agnostic—works with Claude, GPT, Gemini, Llama, Kimi, Ollama, and any tool-calling LLM. Text and vision models may be different providers.
