Clawdcursor
MCP-powered fallback layer that lets AI agents execute tasks through the GUI when APIs, tools, or direct integrations are unavailable. Cross-OS, accessibility-first, local-only.
Last verified:
What is Clawdcursor?
Clawdcursor is an open-source AI desktop agent that gives any AI model 'eyes and hands' on your computer. It connects AI to your desktop by seeing your screen, moving the mouse, typing, clicking, and completing tasks autonomously — essentially doing anything you can do manually. It works with Claude, GPT, Gemini, Llama, and 13+ AI providers including Anthropic, OpenAI, Gemini, Groq, and Ollama.
The tool features an 8-layer OCR-first pipeline that prioritizes cheapest solutions first: OCR reads screen text with bounding boxes, accessibility APIs provide UI automation trees, keyboard shortcuts execute instantly, and vision models are only used as a last resort. It includes 40 desktop control tools available over HTTP on localhost, smart MCP tools (smart_read, smart_click, smart_type, invoke_element), a built-in keyboard shortcuts engine, and a ground-truth verifier that reads actual screen state after every action to prevent false 'done' reports.
Clawdcursor is designed for AI developers, automation enthusiasts, and anyone using AI coding assistants like Claude Code, Cursor, Windsurf, or Zed who wants their AI to have native desktop control. It eliminates the need for separate API integrations — if an app has a UI (Gmail, Slack, Jira, Figma, native apps, legacy software), your AI can use it without OAuth flows, API keys per service, or webhook endpoints.
The tool supports three transport modes: direct integration with Claude Code/Cursor/Windsurf as an MCP server, standalone CLI mode where you type tasks in plain English, and HTTP/REST mode for building custom integrations with OpenAI function-calling format. It runs entirely locally with server binding to localhost only, requires Node.js 20+, and includes three safety tiers (Auto, Preview, Confirm) for different action types.
Clawdcursor pricing
Pricing model: Freemium
Completely free and open source. No paid plans, subscriptions, or tiered pricing. The tool costs $0 for all features including all 40 desktop tools, all 8 pipeline layers, and all 13+ AI provider integrations. You only pay for AI API calls if you use paid cloud providers (Anthropic, OpenAI, Gemini), but free local options like Ollama are supported for $0 cost. Smart tools and Action Router layers execute with zero LLM calls at $0 cost.
Clawdcursor pros
- Open source with no paid tiers or subscription costs
- Works with 13+ AI providers including free options like Ollama
- 40 desktop control tools available over HTTP on localhost
- OCR-first pipeline minimizes costly vision model usage
- Native MCP support for Claude Code, Cursor, Windsurf, Zed
- No separate API keys needed per service — uses your logged-in state
- Ground-truth verifier prevents false success reports
- Skill cache learns action patterns for instant zero-cost execution
- Built-in keyboard shortcuts engine with 30+ shortcuts
- Smart tools enable clicking by element name, not coordinates
- Three safety tiers (Auto, Preview, Confirm) for action control
- Self-healing doctor CLI auto-detects providers and configures pipeline
- ~57ms screenshot capture speed with 1280px HD quality
- Zero LLM calls for simple tasks via Action Router layer
- DPI-safe coordinates for high-DPI displays
- Server binds to localhost only for security
- One-time consent stored locally, no cloud dependencies
Clawdcursor cons
- Requires Node.js 20+ installation prerequisite
- Only supports Windows and macOS (no Linux support mentioned)
- Vision model usage as last resort can be expensive with cloud APIs
- CLI-only interface — no graphical settings UI
- Setup requires running multiple terminal commands
- Limited to 1280px HD screenshots (not full 4K resolution)
- OCR may struggle with non-standard fonts or heavily customized UIs
- Auto-detection of AI providers may fail with custom configurations
- No mobile device support (desktop only)
- Learning curve for understanding 8-layer pipeline architecture
Frequently asked questions about Clawdcursor
What is clawdcursor?
Clawdcursor is an open-source AI desktop agent that gives any AI model eyes and hands on your computer. It connects AI to your desktop by seeing your screen, moving the mouse, typing, clicking, and completing tasks autonomously. It works with Claude, GPT, Gemini, Llama, and 13+ AI providers, eliminating the need for separate API integrations for apps like Gmail, Slack, Jira, and Figma.
Which AI models work with clawdcursor?
Clawdcursor works with 13+ AI providers including Anthropic (Claude), OpenAI (GPT), Gemini, Groq, Ollama, and Llama. It supports any AI model — both cloud providers and local models. The OCR-first pipeline optimizes for cheap text models, with vision models used only as a last resort.
What are the 8 pipeline layers?
The 8-layer pipeline prioritizes cheapest solutions: L0 Browser/CDP for direct DOM interaction, L1 Action Router for pattern-matched tasks and keyboard shortcuts, L1.5 Deterministic Flows for hardcoded sequences, L2 Skill Cache for replaying learned patterns, L2.5 OCR Reasoner (primary layer using OS-level OCR), L2.5b A11y Reasoner as OCR fallback, L3 Computer Use (vision model, last resort), plus a Pre-processor for regex parsing and task decomposition.
How does clawdcursor integrate with Cursor/Claude Code?
For Claude Code/Cursor/Windsurf integration: 1) Run 'clawdcursor consent' once, 2) Add one JSON block to your AI client's config, 3) Desktop tools appear natively in your AI with 40 tools available at $0 extra cost. No extra API key or setup needed. Your AI decides what to do — clawdcursor executes. It plugs in as an MCP server with tools showing up natively in your AI's registry.
What safety tiers does clawdcursor have?
Three safety tiers: Auto for open apps, navigate, read screen (executes immediately); Preview for type text, fill forms (logged before executing); Confirm for send messages, delete files (waits for approval). This provides graded safety from safe reads to destructive actions.
Does clawdcursor send data to the cloud?
No. The server binds to localhost only (port 3847), consent is stored locally in ~/.clawdcursor/consent, and all desktop control happens on your machine. You only send data to AI providers when using their APIs (which you control). The tool itself is open source with no cloud dependencies.
What are the smart MCP tools?
Four new MCP tools: smart_read, smart_click, smart_type, and invoke_element. These enable clicking buttons by name rather than coordinates, requiring no vision model. They work via the 'blind agent' approach using OCR and accessibility APIs instead of screenshots.
How fast is clawdcursor?
Performance benchmarks show: ~57ms for native screenshot capture (1280px HD), ~2s for simple tasks like opening apps and typing via Action Router (zero LLM calls), 100% success rate on tested workflows with self-healing auto-retry. Paint drawing is 78% faster than previous versions, and screenshots are 17× faster after removing VNC dependency.
Can I build custom integrations with clawdcursor?
Yes. Run 'clawdcursor start' to make 40 tools available over HTTP on localhost in OpenAI function-calling format. You can call individual tools or send full tasks, and browse all schemas at GET /tools. Any HTTP client can connect to localhost:3847. This enables building custom AI agents or integrating with existing workflows.