MCP server (for LLM assistants)

WackyTracky can run as an MCP server, letting an LLM assistant (for example Claude Desktop or Cursor) read and manage your tasks for you — "add "buy milk" to my shopping list", "what's on my work list?", "mark task X done".

The MCP server uses the same backend and configuration as the normal server, so it reads and writes the same tasks (e.g. your todo.txt files).

HTTP (Streamable HTTP)

When the WackyTracky HTTP server is running, MCP is available at /mcp on the same origin (for example http://localhost:8080/mcp). This is the recommended transport when your instance is reachable over the network.

Configure your MCP client with the URL and any credentials your deployment requires:

{
  "mcpServers": {
    "wacky-tracky": {
      "url": "http://localhost:8080/mcp"
    }
  }
}

If HTTP authentication is enabled on your instance, add the same Authorization header you would use for the Connect API.

Stdio (local subprocess)

For clients that launch a local MCP subprocess, run the binary with the mcp argument:

wt mcp

This starts the server on standard input/output and waits for an MCP client to connect. You normally don't run this by hand — your LLM client launches it for you (see below). It reads config.yaml from the usual locations (current directory, ../, or /config).

Point your client at the WackyTracky binary with the mcp argument. Most clients use a JSON config like this:

{
  "mcpServers": {
    "wacky-tracky": {
      "command": "wt",
      "args": ["mcp"]
    }
  }
}

If the data directory for the todo.txt backend is not discoverable from where the client launches the process, set it via environment or run from a directory that contains your config.yaml:

{
  "mcpServers": {
    "wacky-tracky": {
      "command": "/usr/local/bin/wt",
      "args": ["mcp"],
      "env": { "DATABASE_DRIVER": "todotxt", "DATABASE_DATABASE": "/path/to/your/todo-dir" }
    }
  }
}

The exact location of this config differs per client (e.g. Claude Desktop's claude_desktop_config.json, or a .cursor/mcp.json for Cursor).

Available tools

Tool	Description
list_lists	List all task lists with their IDs, titles, and item counts.
list_tasks	List the tasks in a list (including subtasks). Requires a list_id.
search_tasks	Search tasks across all lists (#tag, @context, -term to exclude).
create_task	Create a task (todo.txt-style content; optional list_id/parent_task_id).
update_task	Replace a task's content.
complete_task	Mark a task as done.
create_list	Create a new list.

The tools call the same operations as the HTTP API, so behaviour (hide-at-times filtering, todo.txt parsing of #tags/@contexts, etc.) is identical.

Security note

The MCP server has the same access to your tasks as the WackyTracky server itself, with no extra authentication — anything that can launch wt mcp can read and modify your tasks. Only configure it for LLM clients you trust, and treat it like giving that client direct access to your task data.