Connect Claude or ChatGPT to your to-do list with MCP

The Model Context Protocol is the open standard that lets an AI assistant call tools on a server you choose. TaskPocket runs such a server next to your data. It exposes 22 tools over your tasks, projects and areas, the same operations the web app and the iPhone app use, with the same plan rules. Claude and ChatGPT connect to it with the OAuth flow the protocol specifies: the client discovers the server, sends you to a TaskPocket page that names the app and the access it wants, and you approve. Nothing is installed, nothing runs on your machine, and no key is pasted. Clients that configure servers by hand, Claude Code and Cursor among them, can use a personal token instead.

https://api.taskpocket.io/api/public/v1/mcp

This is not a novel idea and is not sold as one. TickTick and Todoist run MCP servers too. What is different here is what stays out of the product: there is no assistant of ours reading your list. The AI is the one you chose, with the access you approved, and it is off until you connect it. That position is laid out on the no-AI page.

Claude calls this a custom connector. Once added it works in claude.ai and in the Claude apps that share your connectors; Anthropic's article lists which plans include custom connectors.

1. 1

   Open the connector settings

   In Claude, open Settings, then Connectors, and choose Add custom connector.

2. 2

   Paste the server URL

   Name it TaskPocket and paste https://api.taskpocket.io/api/public/v1/mcp. Leave the advanced OAuth fields empty: the server registers the client itself.

3. 3

   Approve access

   Click Connect. A TaskPocket page opens in your browser and shows the app's name, the address the access will be sent to, and the scopes it asks for. Sign in if you are not, then Approve. You land back in Claude, connected.

4. 4

   Ask something

   Enable the connector in a chat and try “What is on my plate today?” Claude calls get_context and list_tasks and answers from your real list.

ChatGPT speaks the same protocol and runs the same approval flow. The menu lives behind developer mode, and OpenAI's documentation says which plans can add custom connectors.

1. 1

   Turn on developer mode

   In ChatGPT, open Settings, then Apps & Connectors, then Advanced, and switch on Developer mode.

2. 2

   Create the connector

   Choose Create. Name it TaskPocket, paste https://api.taskpocket.io/api/public/v1/mcp as the server URL, and pick OAuth as the authentication.

3. 3

   Approve the connection in ChatGPT

   ChatGPT sends you to the same TaskPocket consent page. Approve, and the connector's tools register on their own. Enable it in a chat and ask away.

Clients that take a server configuration connect with a personal token. Create one in TaskPocket under Settings → API, choose the scopes it should have, and send it as a bearer header. Claude Code can also run the OAuth flow itself: add the server without a header and run /mcp to sign in.

claude mcp add --transport http taskpocket https://api.taskpocket.io/api/public/v1/mcp \
  --header "Authorization: Bearer tp_pat_your_token"

{
  "mcpServers": {
    "taskpocket": {
      "url": "https://api.taskpocket.io/api/public/v1/mcp",
      "headers": { "Authorization": "Bearer tp_pat_your_token" }
    }
  }
}

A token with the read scope alone makes the assistant a reader: the write tools are not even listed to it. Revoke the token in Settings and access ends on the next request.

Two threads, recorded against the real server. The chips are the tools the assistant called. Hover to pause; the tabs switch threads.

The assistant names what it did and asks before it guesses. When it hits a plan rule, a deadline on a free account for example, it says so instead of pretending.

Every tool runs through the same service layer as the app, so the assistant cannot reach anything you could not reach yourself, and every change shows up on your phone through the normal sync. The scope column is the one on the token or the consent screen.

The same operations are available over REST. The API page documents them, and the interactive reference lists every field.

• It will not think for you. The server has no model and no prompts of its own. Whatever the assistant decides comes from the assistant you connected, with your data as input.
• It will not lift a plan limit. A free account has five projects and two areas, no deadlines, and fourteen days of logbook history, through MCP exactly as in the app. The assistant is told the limits by get_context and gets a plain error if it hits one.
• It will not erase finished work. Deleting a project moves it and its open tasks to the Trash, where you can restore it. Completed tasks stay in the logbook. Trashing a task is reversible; the assistant can restore it.
• It will not share your list with anyone else. Each connection is to one account, and a task id from another account is simply not found.
• It will not run while you are not looking. There are no scheduled jobs, no background agents and no webhooks. The assistant only acts inside a conversation you are having.

The consent screen shows two things worth reading: the app's name, which the app chose, and the address the access will be sent to, which the app cannot fake. Access tokens live for an hour and renew themselves; a grant lives at most a year before it asks you again. Settings → API → Connected apps lists every grant, and Disconnect ends one at once. A read-only grant cannot be talked into editing: the write tools are hidden from it and refused if called by name. If an account is suspended, every connection becomes read-only, the same as the app.

What the assistant does with what it reads is between you and its vendor. TaskPocket sends your data to the client you connected and to nobody else.