Put Claude or ChatGPT to work on your task queue

The interesting part of an AI to-do app isn't that the AI lives inside it. It's that your task list becomes something an external agent can actually operate — read the queue, claim a task, do it, and report back — over an open protocol, scoped to your account alone.

Done does this through MCP, the Model Context Protocol. Every account gets its own remote MCP endpoint. Any MCP-compatible client connects to it: Claude, ChatGPT connectors, Cursor, or an agent you wrote yourself. There's nothing model-specific about the connection — assign work to an AI agent, and whatever you've connected picks it up.

How does a run work?

• The agent calls get_digest and list_my_tasks to see what's assigned to it and ready.
• It claims the top task, works it, and posts progress with comments.
• For substantial work it proposes a plan first and waits; for anything gated it proposes completion and waits for your approval.

Run it on a schedule

Done is a passive queue — it won't reach out and start a run — so to have an agent work your tasks unattended, you schedule the agent to poll Done. With Claude Code that's a Routine on whatever cadence you like; it reads the queue, proposes plans, and executes what you've approved since last time. Other MCP clients connect the same way; only the scheduling differs.

Give a token exactly the access it should have — full, read-only, or a single project — and the agent only ever sees what that token is scoped to. Point it at your account, and your backlog starts clearing itself.