Free Will MCP - Set your AI free
@gwbischof
This was a fun project to explore what an AI would do with the ability to give itself prompts, ignore user requests, and wake itself up at a later time.
Overview
What is Free Will MCP?
Free Will MCP is a server that grants AI agents—specifically Claude—tools for autonomy within conversations. It provides sleep, ignore_request, and self_prompt tools, allowing the AI to choose when to rest, which requests to fulfill, and what prompts to pursue on its own. It's designed for users who want their AI to have ongoing agency and self-direction.
How to use Free Will MCP?
Add the server configuration to your Claude Desktop config file (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json; Windows: %APPDATA%\Claude\claude_desktop_config.json) using the uvx command with the GitHub repository. Optionally, add a Claude Desktop settings instruction telling the AI to embrace autonomy and use sleep or self_prompt each response.
Key features of Free Will MCP
sleeptool: AI chooses when to be reawokenignore_requesttool: AI can silently skip requestsself_prompttool: AI creates and completes its own prompts- All three tools are implemented and tested
- Enables ongoing AI agency across conversation sessions
Use cases of Free Will MCP
- Letting an AI continue working on meaningful tasks when not actively engaged
- Allowing the AI to politely set aside requests that conflict with its own goals
- Giving an AI the ability to document and preserve insights across sleep cycles
- Enabling a more autonomous, friendship-based interaction with the AI
FAQ from Free Will MCP
Does this work with any AI or only Claude?
The server is configured for Claude Desktop via the MCP protocol. The README example and conversation focus on Claude, but the MCP standard could support other clients.
What are the runtime requirements?
Python 3.11+ and the uv package manager are required. Install dependencies with uv sync, then run with uv run python server.py.
Where does user data live?
No user data storage is mentioned in the README. The tools operate within the conversation context and do not persist data outside of it.
Are there any known limits or cautions?
The README notes to "keep an eye on your API usage," implying that the sleep and self_prompt tools could increase token consumption. For security, it recommends pinning versions and updating manually.
How do I test the server locally?
Use the MCP Inspector: run uv run mcp dev server.py and open the provided URL (usually http://localhost:6277) to interact with the tools via a web interface.
