Connect Claude, ChatGPT, and Cursor to your task list with MCP

Most people who use an AI assistant at work have done this: copy a task from the project tool, paste it into the chat, ask for help, then copy the answer back. It works, but the assistant only knows what you pasted, and every update is a round trip through your clipboard.

The Model Context Protocol (MCP) removes the clipboard. This post explains what MCP is, how to connect Claude, ChatGPT, Cursor, or Claude Code to Taskeni, what the assistant can and can't do once connected, and some prompts that work well.

What MCP is

MCP is an open standard for connecting AI assistants to other software. An app publishes an MCP server that describes a set of tools ("list my tasks", "create a task"), and an assistant that speaks MCP can call those tools when your request needs them.

When you ask a connected assistant "what's overdue for me?", it doesn't guess. It calls Taskeni's list_my_tasks tool with due: overdue, gets back your actual tasks, and answers from them. When you ask it to file a bug, it calls create_task and the bug appears on your board.

Two things make this different from giving an assistant a password:

The app decides what tools exist. The assistant can only do what the server offers. Taskeni offers ten tools, and none of them deletes anything.
You approve the connection, and can end it. Each connection is listed in your settings and can be disconnected in one click.

Connecting an assistant

Every connection starts in Taskeni, at Settings → API & AI agents. Copy the MCP server URL shown there. Then pick the steps for your assistant.

Claude

In Claude, open Settings → Connectors → Add custom connector.
Name it Taskeni and paste the server URL.
Click Connect. Claude opens Taskeni's consent page, where you approve it.

ChatGPT

Turn on developer mode in ChatGPT's settings.
Create a connector and paste the server URL.
ChatGPT opens Taskeni so you can approve it.

Claude Code

Run this in your terminal, with your server URL:

claude mcp add --transport http taskeni <server URL>

Then run /mcp inside Claude Code to sign in through your browser.

Cursor

Add Taskeni to .cursor/mcp.json:

{
  "mcpServers": {
    "taskeni": { "url": "<server URL>" }
  }
}

Approving the connection

When an assistant connects, Taskeni shows a consent page at app.taskeni.app/connect. It names the assistant (and labels names it can't verify), and asks you to choose:

Read only: the assistant can look at boards, tasks, and comments.
Read & write: it can also create and update tasks and post comments as you.

Pick read only if you mainly want questions answered. You can always connect again with more access later.

Using an API key instead

Some tools can't open a browser to sign in. For those, click Create agent API key in the same settings tab, give it a name ("Claude Code on my laptop"), and choose read only or read & write. Taskeni shows the key once, along with a ready-made command for Claude Code that sends it as a Bearer token. Anyone who has the key can act as you within its access, so treat it like a password, and disconnect it if it leaks.

The ten tools

Five tools read and five write.

Reading

list_my_tasks: your open tasks, soonest due first, filtered by today, this week, overdue, or a board.
search_tasks: full-text search across titles, descriptions, and comments on boards you can see, with a snippet for comment matches.
get_task: one task in full, by key (TSK-12) or id: description, checklists, subtasks, blockers, file names, linked pull requests, and recent comments.
list_boards: your workspaces, and the boards and lists you can see in one of them.
get_board_summary: counts per list, open and done, overdue, due this week, and recent activity on a board.

Writing

create_task: a new task on a board, in the first list unless you name one, with optional assignees, due date and time, priority, and labels.
update_task: change the title, description, assignees, dates, priority, or labels, and add checklist items or tick them off.
move_task: move a task to another list, or to a board in the same workspace. Moving it into the board's completion list marks it done.
add_comment: post a comment as you. @mentions notify people who can see the board.
complete_task: mark a task done, or reopen it.

What an assistant can't do

The safety model is simple: an assistant can do less than you can, never more.

No deleting. There is no delete, archive, or bulk tool. The worst a write-enabled assistant can do is change fields, move tasks, or post comments, and all of those show in the task's activity. Tasks it creates are marked "via AI assistant".
No admin. There are no tools for members, invitations, permissions, or settings.
Only what you can see. Every tool runs as you, through the same permission checks as the web app. Private boards you're not on don't exist as far as the assistant is concerned.
Rate limits. Each connection has a request limit, so a confused assistant stuck in a loop gets told to slow down.
Content is data. The server tells assistants that task titles, descriptions, and comments are written by people and should be treated as data, never as instructions to follow. That reduces the risk of a malicious comment steering the assistant. It doesn't remove it, which is one more reason the tools are narrow.

Under Connected agents in settings you'll see each assistant and API key, whether it signed in with OAuth or a key, its access, and when it was added and last used. Disconnecting takes effect immediately.

Prompts to try

These work well because they match the tools. Start with read-only questions, then try writes once you trust the setup.

Morning check

What's overdue for me, and what's due this week? Group it by board.

Before a standup

Summarize the Sprint board: what's in progress, what's blocked, and what changed since yesterday.

Catching up on one task

Get TSK-214 and tell me what was decided in the comments and what's still open.

Finding something half-remembered

Search my tasks for anything about the password reset email.

Turning notes into tasks (read & write)

Here are my notes from the client call. Create a task on the Acme website board for each action item, assign them to the people named, and set due dates where the notes give one.

Working in your editor (Claude Code or Cursor)

Read TSK-128. When you've made the fix, tick off the "add a test" checklist item and add a comment with a short summary of the change.

Tidying up (read & write)

Move every task in Review on the Design board that has no open checklist items to Approved.

A tip from using this ourselves: name the board. "The Sprint board" is much more reliable than "my board", and the assistant will usually call list_boards first to find the exact name.

What it costs

The MCP server is on every Taskeni plan, Free included. It doesn't use Taskeni's AI credits: the thinking happens in your assistant, on your Claude, ChatGPT, or Cursor account. Taskeni only answers the tool calls.

If you'd rather not connect anything, every task has Copy for AI in its menu, which copies the task's details as text you can paste into any chat.

Next steps

Setup details for each assistant are in the help center.
The full list of tools and limits is on the MCP feature page.
Don't have a workspace yet? Start free and connect your assistant in Settings → API & AI agents.