Codex task manager: give OpenAI Codex a task queue over MCP
How to connect OpenAI Codex (CLI, IDE extension or ChatGPT desktop app) to Done over MCP, what it can do there, and why it still can't close a gated task on its own.

By Alejandro Rioja5 min read
To give Codex a task queue, add your task manager as a remote MCP server in Codex's config. With Done that is two commands: codex mcp add done --url https://trydone.app/mcp, then codex mcp login done. After that Codex can pull the next task from your queue, work it, comment on it and propose completion, and anything you haven't marked auto-OK still waits for your approval before it counts as done.
How do I connect Codex to Done?
The Codex CLI, the Codex IDE extension and Codex in the ChatGPT desktop app share one configuration file (~/.codex/config.toml), so you set Done up once. There are two ways in:
- Sign in with OAuth: run codex mcp add done --url https://trydone.app/mcp, then codex mcp login done. Approve access in the browser window that opens. Codex registers itself with Done and receives the sign-in on 127.0.0.1, so there is no token stored in your files.
- Use a token: generate an MCP token in Done under Settings → Agent access, export it as DONE_MCP_TOKEN, and add a [mcp_servers.done] table with url = "https://trydone.app/mcp" and bearer_token_env_var = "DONE_MCP_TOKEN". This is the route for a remote or headless machine, where the browser sign-in can't reach Codex's local callback.
- ChatGPT on the web doesn't read Codex's config file. It adds Done as a custom MCP server instead; see the ChatGPT post below.
What can Codex do once it's connected?
Codex gets the same tools any MCP client gets. next_task atomically picks and claims the highest-priority task that is ready, so two agents never take the same one. get_task loads the notes, subtasks, blockers and activity. comment_on_task posts progress where you'll read it, beside the task. When the work is done, complete_task closes a task marked auto-OK, and propose_completion posts a summary and parks every other task in pending approval.
Can Codex close a task without my approval?
Only if you marked that task auto-OK. Whether a task may close without a human is your grant, not the agent's: an agent can't set agent_ok when it creates or edits a task, and a task that isn't agent_ok can only reach done through your approval. Codex can do the work and propose it. It can't approve itself.
Can I limit what Codex can touch?
Yes, with the token route. A token is per account, never shared, and can be read-only or limited to one project, so a Codex session in one repo needn't see the rest of your board. Done doesn't launch Codex or run it on a schedule. Whatever starts Codex, you or your own scheduler, is what makes it check the queue; the schedule guide covers the loop.
Frequently asked questions
How do I add an MCP server to Codex?
Run codex mcp add <name> --url <server-url>, then codex mcp login <name> for OAuth, or add an [mcp_servers.<name>] table to ~/.codex/config.toml by hand. For Done the URL is https://trydone.app/mcp.
Does the Codex setup work in the IDE extension and the ChatGPT desktop app?
Yes. The Codex CLI, the IDE extension and Codex in the ChatGPT desktop app read the same config file, so one setup covers all three. ChatGPT on the web doesn't, and uses a custom MCP server instead.
Why does the Codex sign-in hang on a remote machine?
Codex waits for the sign-in on 127.0.0.1 of the machine it runs on, which your browser can't reach over SSH. Use the token setup there.