Managing tasks from ChatGPT, Claude or Gemini: what needs to connect

To manage a real task list from an AI assistant, the assistant needs an authorised connection to the service that stores the tasks. Asking ChatGPT, Claude or Gemini to “plan my week” is useful, but on its own it produces text, not a change to your to-do list. This guide explains what has to connect, how the Model Context Protocol (MCP) fits in, and what to check before you give an assistant access to your tasks, habits or journal.

A plan in a chat is not a saved task

When you ask an assistant to plan your week, it writes a plan. That plan lives in the conversation. If you start a new chat, open your task app on your phone, or tick something off, the plan doesn’t know.

For the assistant to change your list, by adding a task, moving a date or marking something done, three things need to be in place:

1. A service that keeps the record. Your task manager, which stores tasks and their state.
2. A tool the assistant can call. A defined action like “create task” with clear inputs.
3. Permission. Proof that you, specifically, allowed this assistant to act on your account.

The assistant handles the request; the service keeps the record

Take “Add the proposal to Friday’s tasks.” With a connection in place:

1. The assistant works out what you want: a task called “the proposal”, due Friday.
2. It calls the task service’s tool with those details.
3. The service checks permissions, validates the date, saves the task and returns the confirmed result.
4. The assistant shows you what was saved.

That separation is what makes it reliable. The task exists in your task manager, not in the chat, so you can complete it from your phone later, move it in a calendar view, or ask a different assistant about it tomorrow.

What MCP contributes

The Model Context Protocol (MCP) is an open standard for connecting AI assistants to tools supplied by other services. A service publishes an MCP server listing its tools, such as get_schedule, create_task or set_done, and any compatible assistant can discover and call them.

MCP is the plumbing, not the product. It doesn’t give you an account, decide your priorities, or guarantee every assistant supports the same features. Authentication, permissions and each app’s own setup still matter. But it means one server can work with Claude, ChatGPT, Claude Code, Cursor and other MCP clients rather than needing a separate integration for each.

What to check before you connect

Not every connection is built the same way. Before you connect an assistant to your tasks or journal, look for:

• OAuth sign-in. You should approve the connection on the service’s own site, and the assistant should never see your password. Be wary of setups that ask you to paste API keys or passwords into a config file.
• A clear list of actions. The consent screen should say exactly what the assistant can do: read tasks, create tasks, read journal entries, and so on.
• Limited scope. A fixed set of actions on your own data, not raw database access.
• Confirmed results. The assistant should report what was actually saved, not just say “done”.
• Undo. Assistants misread things. You want an easy way to reverse a change.
• Disconnect. One place to see connected assistants and remove access.

Where each assistant stands

Support varies by app and changes often, so treat this as a snapshot:

• Claude supports custom remote MCP connectors on the web and desktop, and Claude Code can add MCP servers from the command line.
• ChatGPT supports custom connectors built on MCP; depending on your plan, you may need to turn on developer mode to add your own.
• Gemini has its own developer flow for connected apps. Signing in to a service with Google, or connecting Google Calendar, does not by itself connect that service to Gemini.

Whatever the assistant, the same principle holds: the service keeps the record, and the assistant asks it to make changes on your behalf.

Can a task list be more than text?

The MCP Apps extension lets compatible assistants show a small interactive view supplied by the tool server, for example a task list with checkboxes inside the conversation, instead of only a paragraph of text. The assistant still controls the surrounding experience, and each host supports different capabilities. See the MCP Apps overview for how it works.

What should stay easy without AI

Some actions are faster without a conversation. Completing one task, correcting a habit check-in or editing a journal entry should be one tap in an app you can open directly. An assistant is great for capturing, planning and reviewing; it shouldn’t be the only way in.

How this works in Divitica

Divitica is a task manager, habit tracker and journal built for this model. Its MCP server at https://divitica.com/mcp gives your assistant a fixed set of actions:

• Tasks: list, search, create, update, complete and delete.
• Habits: list with the last 7 days of history, create, and check in for today or up to 60 days back.
• Journal: read up to 31 days of entries, and save new entries in your own words with an optional 1–5 mood.
• Undo: reverse any change the assistant made.

You sign in and approve on divitica.com using OAuth 2.1 with PKCE. Access tokens expire after an hour, tokens are stored only as hashes, and you can disconnect any assistant from Settings → Connected Assistants. Divitica works with Claude, ChatGPT and other MCP-compatible tools today; we’ll list Gemini once Divitica is available there.

Everything an assistant saves appears in Divitica’s Day, Week and Month views, where the direct controls stay one tap away. See the setup guide, the example prompts, or explore the task manager, habit tracker and journal.