ChatGPT task manager: connect ChatGPT to your to-do list over MCP
By Alejandro Rioja
A "ChatGPT task manager" is a task list ChatGPT can actually operate on — not just discuss. Done gets there the same way it does for any other model: an MCP endpoint per account, with the same tools and the same guardrails, regardless of which client is on the other end.
Connecting ChatGPT over MCP
ChatGPT's connector settings are OpenAI's surface, not Done's, so the exact menu names shift as the product updates. The shape stays the same either way: add a custom connector, point it at https://trydone.app/mcp, and authorize it. For a plan or client that doesn't speak OAuth, generate a bearer token from Done's Account panel and use that instead — the same fallback Claude Code or a custom script would use.
- OAuth-capable ChatGPT connectors: add Done as a custom connector by URL, then sign in and approve — no token to copy or store.
- A bearer-token fallback: generate a per-account token from the Account panel and set it in whatever config your ChatGPT-based client exposes.
Same tools, same account, no lock-in
This is the point of building on MCP rather than a bespoke ChatGPT plugin: Done exposes one set of tools, and any MCP-compatible client gets the same ones. Nothing about your tasks, projects, or structure is Claude-shaped or ChatGPT-shaped — switch models, or run both against the same account at once, and neither one needs your task list rebuilt around it.
What actually differs between Claude and ChatGPT here
The underlying tools and data are identical. What differs is entirely on the client side — how each product schedules sessions, how long a context window it brings to a task, how it surfaces a connector in its own settings. None of that is something Done controls or should try to paper over with a fabricated comparison; it's genuinely just "ask the client," not "ask Done."
One thing that's true of every client, ChatGPT included: nothing reaches out and starts a run on its own. Done is a passive queue. Whatever you connect needs its own scheduling — a recurring job, a routine, or you opening it and asking it to check the queue — to actually clear work unattended.
A typical ChatGPT run
The flow is the same regardless of model: pull the digest for what's overdue and due today, claim the top task from the queue, load its full context, do the work, and either complete it outright or propose completion and wait if the task isn't auto-OK. Whichever model is attached, that sequence — and the approval gate at the end of it — doesn't change.
Guardrails apply the same way
Scoped tokens, the approval gate, and run budgets are all enforced by Done's server, not by the client. A ChatGPT connection gets exactly the same protection a Claude connection does: read-only or single-project tokens if you want them, gated tasks that can't self-approve, and every session logged as an auditable run.
Frequently asked questions
Can ChatGPT create and complete tasks in Done?
Yes. Done is MCP-native, so any MCP-compatible client — ChatGPT connectors included — can connect to your account, pull its queue, claim a task, and complete or propose completion on it, the same as Claude or a custom agent.
Is connecting ChatGPT different from connecting Claude?
Not on Done's side — same MCP endpoint, same tools, same account scoping. What differs is entirely in each client's own settings UI and scheduling model, which Done doesn't control.
Do I need a paid ChatGPT plan to connect it to Done?
That depends on which ChatGPT surface supports custom MCP connectors or OAuth-based tools, which is OpenAI's call, not Done's — check what your plan exposes. A bearer token is the fallback for any client that can't do OAuth.