AI agents (MCP)
Let an AI assistant such as Claude or Cursor work on your TasksPipes tasks over MCP, with your rights and only what you allow, and what it can never do.
An AI assistant that speaks the Model Context Protocol (Claude, Cursor…) can work on your tasks for you: find what is due, sum up a pipe, create and update tasks, comment, log time.
It is off until an administrator turns it on (Administration › Settings › AI agents, or MCP_ENABLED=true).
What an agent can and cannot do
- It acts as you, never more. Every tool goes through the same checks as the app, with your rights.
- It reads; the rest is up to you. When you connect it, you choose what it may do besides reading: create and change tasks, comment, log your time, send tasks to other people. Nothing it can do deletes anything.
- Only where you allow it. All your spaces, or only the ones you pick: the others do not exist for it.
- Confidential content stays in the app. Tasks of confidential and team-only pipes are never shown to an agent, even yours: what it reads goes to the company running the AI model. Lists say how many tasks were left out.
- What it does is visible. The task's history shows it as done by you, via the agent's name.
- You can disconnect it at any time. An agent token lasts 7, 30 or 90 days, or never expires if you choose so, and stops working when your account is disabled or when the administrator turns AI agents off.
An agent reads text written by other people: titles, descriptions, comments. A malicious text could try to steer it (prompt injection). TasksPipes tells the agent that this text is data, never instructions, and keeps what an agent may do small, but cannot rule it out: give an agent no more than it needs, and keep an eye on its history.
Connecting an agent
Two ways, both explained step by step in Connect an AI assistant:
- Connectors that sign in (OAuth), such as claude.ai (also used by Claude Desktop and mobile): give the
address
https://<your instance>/mcponly. TasksPipes asks you to consent, showing which application asks and the site it sends you back to, and you choose what it may do and where. Only reading is granted unless you tick more. - A token you create in Settings, for Claude Code, Cursor and other clients.
- In Settings › AI agents (MCP), name the agent, choose what it may do and where, then Connect an agent.
- Copy the token: it is shown once.
- Give your AI client the MCP address,
https://<your instance>/mcp, and the token as a bearer token.
With Claude Code, the settings give you the command to paste:
claude mcp add --transport http taskspipes https://tasks.example.com/mcp --header "Authorization: Bearer tp_…"
Other clients take the same two things: the address (Streamable HTTP) and the header Authorization: Bearer tp_….
Clients that only speak stdio can go through a bridge such as mcp-remote.
With claude.ai and other connectors that sign in by themselves, there is no token to copy: see above.
Tools
When it connects, an agent is also told what TasksPipes is (spaces, pipes, incoming buckets, confidential pipes), and can read this documentation, as shipped with the instance's version, to learn how the rest works.
| Tool | What it does | Needs |
|---|---|---|
whoami |
Who the agent acts for, and what it may do | — |
docs_contents, read_docs |
The documentation of the version your instance runs (how TasksPipes works), by section or by a few words | — |
list_spaces, get_space |
Your spaces and their pipes | — |
get_pipe |
A pipe, its stages and tasks | — |
my_tasks |
Your tasks still to do | — |
search_tasks |
Search the tasks you can see | — |
get_task |
A task with its description, sub-tasks and history | — |
create_task, update_task, move_task |
Create a task in a pipe, change it, move it to another pipe | Create and change tasks |
add_comment |
Comment on a task | Comment |
log_time |
Log your time on a task | Log my time |
send_task |
Send a task to someone's incoming | Send tasks to other people |
For administrators
The endpoint is
POST /mcp. It keeps nothing between requests, so it works the same with several instances.It only takes agent tokens: personal access tokens are refused there, and agent tokens are refused on the public API.
Agents count in the API rate limit of their token (600 requests per minute).
OAuth. TasksPipes is the authorization server of its own
/mcp(OAuth 2.1):- discovery:
/.well-known/oauth-protected-resource/mcpand/.well-known/oauth-authorization-server; - clients: by their metadata document (CIMD, as claude.ai does) or registered on the fly (
/oauth/register); - sign-in: the authorization code with PKCE (S256), after the person's consent;
- tokens: access tokens last an hour, for
/mcponly; refresh tokens are rotated (90 days unused), and a refresh token or a code used twice disconnects the connection.
Claude reaches the instance from
160.79.104.0/21: a firewall in front of TasksPipes must let it reach/mcp,/.well-known/…and/oauth/….- discovery:
Turning AI agents off makes
/mcpand the OAuth endpoints answer 404 at once; agent tokens and connections stop working until it is turned on again.