Capturing tasks with AI assistants: MCP explained

You are fixing a bug with Claude Code and notice that the documentation of another module is out of date. That used to mean: switch windows, type a task, switch back, lose your train of thought. With an AI assistant that knows your task list, you just say “add a task for that, due Friday”. An open standard called MCP makes this possible.

What MCP is

The Model Context Protocol (MCP) describes how an AI assistant talks to other software. Anthropic published it as an open standard in November 2024 (announcement, specification). Since then many clients have adopted it, among them Claude Code, Claude Desktop and Cursor.

The idea is simple: an MCP server offers tools, such as “list tasks” or “create a task”. The client, your assistant, sees the list of these tools with a short description and calls one when it fits your request. The assistant does not get free access to a database. It gets exactly the tools the server offers, and only with the rights of your access token.

What an assistant can do with your tasks

tasks has its own MCP server. It offers these tools:

Tool	What it does	Needs
list_workspaces	lists your workspaces	read
list_tasks, get_task	lists tasks and shows one in detail	read
list_taxonomy	lists the categories, tags and priority labels of a workspace	read
quick_add	creates a task or, with parent_id, a subtask from one line, with the same quick add as in the browser	write
create_task, update_task	creates a task from individual fields or changes one	write
set_status	changes the status, for example to done	write
set_parent	makes a task a subtask or a top-level task again, within one workspace	write
add_comment	adds a comment to a task	write
create_tag, create_category	adds a tag or a category (by its path, e.g. Home/Garden); a category needs the owner or admin role	write

One tool is missing on purpose: delete. No assistant can delete a task through MCP. If you really want to delete one, do it in the app or with tsk rm in the terminal.

Which permissions an assistant needs

The most important rule for AI assistants is as few rights as possible. Not because assistants are malicious, but because they process text they did not write: a web page, an email, a file in the repository. If that text contains a hidden instruction, it can change what the assistant does. OWASP lists this prompt injection first among the risks of language model applications and recommends limiting a model’s privileges to what it needs (OWASP LLM01:2025).

For your tasks, that means:

• Read-only is often enough. If you want to ask Claude what is due this week, it needs no write access. A token with read access only (tasks:read) hides the write tools completely. The assistant does not even see them.
• Allow writing on purpose. If the assistant should create tasks, add tasks:write to the token. Deleting stays impossible either way.
• Limit it to workspaces. A token can be restricted to selected workspaces. The assistant in your company repository then sees your work tasks, but not the family shopping list.
• One token per tool. One token for Claude Code, one for Cursor: you can see which one was used last and revoke one without touching the others. Every token also expires.

Review before saving

An assistant can misread a request: the wrong date, the wrong workspace. Many clients, Claude Code among them, therefore ask before calling a tool until you explicitly allow it. Keep that prompt on for the write tools, at least at the start. Every task created through MCP also lands in your list like any other, so a wrong one is quickly fixed.

If you do not want an assistant

MCP is one way among several. If you live in the terminal, tsk add captures a task in one line with no AI in between. Every channel goes through the same quick add; more on the page Terminal task manager.

How to set it up

Setup takes a few minutes: create a token in the settings, choose its rights, add the server to Claude Code, Claude Desktop or Cursor. The steps for each client are in the guide Set up MCP. What else the MCP server can do is on the page Capture tasks straight from Claude.

Conclusion

An AI assistant is a good place to capture tasks, because you are already working there. Give it no more than it needs: read access for questions, write access only when it should create tasks, and only for the workspaces it has to see.