Quickstart

You’ll need:

• Node.js 20+
• An Anthropic API key
• An e2b API key (free tier is plenty)

1. Add Patch to your MCP host

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "patch-cat": {
      "command": "npx",
      "args": ["-y", "@patch-cat/mcp"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "E2B_API_KEY": "e2b_..."
      }
    }
  }
}

Cursor

Edit ~/.cursor/mcp.json (or per-project .cursor/mcp.json):

{
  "mcpServers": {
    "patch-cat": {
      "command": "npx",
      "args": ["-y", "@patch-cat/mcp"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "E2B_API_KEY": "e2b_..."
      }
    }
  }
}

Claude Code

claude mcp add patch-cat --scope user \
  --env ANTHROPIC_API_KEY=sk-ant-... \
  --env E2B_API_KEY=e2b_... \
  -- npx -y @patch-cat/mcp

Windsurf

Same JSON shape as Claude Desktop, in Windsurf’s MCP config. Path varies by version; consult Windsurf’s docs.

2. Restart your host

The host loads MCP servers at startup. After editing config, fully quit and relaunch Claude Desktop (or whichever host you’re using).

3. Try a task that needs a tool

In the host:

“Fetch the top 5 stories from Hacker News and rank them by points.”

The host will call patch_generate_tool with that description. Patch:

1. Asks Claude Opus for a Python tool in the manifest format.
2. Runs py_compile on the result inside an e2b sandbox.
3. Saves the file to your local toolbox.
4. Notifies the host (via notifications/tools/list_changed) that a new tool is available.

The new tool appears in your host’s tool list within ~1 second. The host then calls it directly. Patch spins up a sandbox, runs pip install, executes the tool, parses the JSON output, returns it.

Quit and reopen the host. Ask the same question again. Patch uses the existing tool — no LLM call, no regeneration. The toolbox is permanent.

What’s at the local toolbox path

Resolved via env-paths:

Platform	Path
macOS	~/Library/Application Support/patch-cat/
Linux	~/.config/patch-cat/
Windows	%APPDATA%\patch-cat\

Inside:

patch-cat/
├── tools/             one .py file per tool, manifest frontmatter + source
├── index.json         tool name → version, file path, last-used timestamp
├── runs/              forensic audit log — one .json per execution
│   └── blobs/         content-addressed stdout/stderr by SHA-256
└── config.json        registry URL, contribute opt-in, etc.

Enabling the registry contribute path

Default behavior: Patch reads from the public registry but doesn’t contribute. To opt in:

In your host AI, ask:

“Use patch_auth_register to authorize me to contribute tools to the registry.”

The host calls patch_auth_register({ provider: "github" }). Patch starts a local OAuth listener, hands back a URL. Click the URL in your browser, complete GitHub OAuth, return. The contribute_token lands in your local config; subsequent tool generations get pushed back to the registry.

You can revoke at any time via the GitHub OAuth apps page.

What if I want to inspect what just ran?

Ask the host AI: “Use patch_list_runs to show me the recent tool executions.”

Each run has a run_id. Pass it to patch_replay to re-run the exact same source against the same inputs in a fresh sandbox and confirm the output matches. See Architecture → Replay for what replay actually proves.