Antigravity Learn MCP Server (andy)

A Model Context Protocol (MCP) server that mimics the Google Antigravity TUI slash command /learn for external agent harnesses like Cursor and Claude Code (Claude CLI).

It allows agents to inspect their own history, propose new rules or skills, and automatically write/update configurations under .cursorrules, .agents/AGENTS.md, or globally in ~/.gemini/config/.

Features

• Prompt (learn): A guide that walks the LLM through the learn workflow (analyzing history, classifying rules/skills, proposing, and applying).

• Tools:

  • list_customizations: Lists workspace and global rules and skills.

  • read_customization: Reads any rule or skill file.

  • write_proposal: Writes a draft learning_proposal.md for user review.

  • apply_customization: Persists the rule or skill to disk.

• Resources:

  • rules://workspace: Exposes local .cursorrules and .agents/AGENTS.md rules.

  • rules://global: Exposes global ~/.gemini/config/AGENTS.md rules.

  • skills://workspace/list: Summarizes local workspace skills.

  • skills://global/list: Summarizes global skills.

Related MCP server: OpenClaw Memory

Install

You need uv (a fast Python package manager):

brew install uv

Option 1 — Assisted setup (recommended)

Copy this prompt and paste it into any AI assistant (Claude Code, Cursor, etc.). It will walk you through the rest:

I want to install andy learn (https://github.com/timtty-sinch/andy_learn_tool) to run the /learn workflow in Claude Code or Cursor. Please guide me step-by-step, including installing it and running the setup wizard.

You can also print this prompt any time with andy prompt.

Option 2 — Manual install

uv tool install git+https://github.com/timtty-sinch/andy_learn_tool
andy setup

andy setup is an interactive wizard that asks which agent you use (Claude Code, Cursor, or both) and registers the server for you:

• Claude Code → runs claude mcp add andy --scope user -- andy serve

• Cursor → writes an andy entry into .cursor/mcp.json (workspace or global)

Verify Claude Code registration with claude mcp list. Restart your agent afterward.

Manual registration (without the wizard)

# Claude Code
claude mcp add andy --scope user -- andy serve

For Cursor, add to .cursor/mcp.json:

{
  "mcpServers": {
    "andy": { "command": "andy", "args": ["serve"] }
  }
}

Running from a source checkout (dev)

Without installing, run the script directly — andy setup will register uv run against the script path automatically:

uv run andy_learn_mcp.py setup

How to Trigger the /learn Workflow

In Claude Code

Type the following in your Claude prompt:

Use the learn prompt to analyze this session and persist any rules or skills.

Claude will:

1. Load the learn prompt instructions.

2. Scan the current terminal session for corrections/successes.

3. List existing customizations to prevent duplicates.

4. Call write_proposal to write a draft learning_proposal.md to your workspace root.

5. Present the draft in chat and ask for your approval.

6. Once you say "yes", call apply_customization to save it.

In Cursor Chat

1. In the chat input, type @ and select Prompts (if available) or simply tell the model:

   Run the andy prompt to extract and save lessons from our recent work.

2. The agent will read its context, write learning_proposal.md for your review, and wait for your confirmation to save the files.