> ## Documentation Index
> Fetch the complete documentation index at: https://docs.jacobpevans.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Task management

> A self-hosted task and project tracker that autonomous agents read and write through an MCP server, so agent and operator share one board.

<Note>
  Status: **live.** The tracker and its MCP server are deployed; agents read and
  write it today.
</Note>

> The agent and the operator plan against the same board. Neither keeps a private
> to-do list the other can't see.

Autonomous agents need somewhere durable to record what they are doing and what is
left to do — a store that survives a restart, a new session, or a handoff to a
person. A self-hosted task tracker fills that role: projects, tasks, labels, and
status, shared by every agent and every human on the homelab.

## What the agent can do

Agents reach the tracker through an [MCP](/ai-development/ai-orchestration-stack)
server that exposes project and task operations — create, read, update, close,
and organize — as tools. The agent plans work as tasks, records progress, and
closes items it finishes, the same way a person would from the web UI.

| Capability                             | Who uses it        |
| -------------------------------------- | ------------------ |
| Create and organize projects and tasks | Agent and operator |
| Update status, labels, priority        | Agent and operator |
| List work per project                  | Agent and operator |
| Shared board of record                 | Both, one view     |

## Credentials, not prompts

The MCP server authenticates with a **scoped service-account token** (a read-write
tier, distinct from any human account) that is injected at process launch by the
[secrets path for agents](/autonomous-agents/secrets) — never typed, never stored
in the clone. An operator opts a host into the integration deliberately; it is off
by default everywhere else.

## Known limits

Listing every task across all projects in one call is unreliable — agents list per
project instead. This is a client-side quirk, not a data limit.