MCP for project management: AI agents in your tracker

Most integrations between AI assistants and project management tools top out at “read a value and paste it somewhere.” The agent can summarize your board, but it cannot move a card, update a deal stage, or acknowledge a customer. The Model Context Protocol (MCP) changes that boundary: it is an open protocol that lets an AI agent act on your data through the same tools your team uses every day, rather than a parallel set of read-only endpoints.

What MCP is, and why it matters here

MCP is a protocol, not a product. It defines how an AI assistant discovers a tool, authenticates to it, and calls it — and crucially, how it reads back the result. A server built on MCP advertises 70+ tools the same way your team does: create issue, list tickets, update a deal, append to a page. The agent sees the same shape the human would see, so the handoff between agent and human is a natural pause, not a translation layer.

That matters for project management because the work is collaborative and stateful. An agent that can only read your board cannot reduce the noise in your inbox; an agent that can create, assign, and close the same issues you can, can.

Where agents earn the cost

The practical uses are smaller than the hype suggests, and that is the point. Three that teams actually keep running:

• Triage an inbox. An agent reads support mail, drafts a reply that a human reviews, and creates a linked issue when the thread needs engineering input — all in the project’s own states.
• Release notes from done issues. At the end of a cycle, an agent scans issues in the Done state, drafts a changelog grouped by component, and opens a page for the release.
• “What is blocked?” queries. An agent can answer “which deals are stuck waiting on legal?” or “which issues have been in review for over three days?” faster than building a saved view, because it can ask follow-up questions about the same data.

How Atur does it

Atur ships a built-in MCP server with 70+ tools covering issues, tickets, CRM, time entries, and pages. Authentication is a single API key, and scoped keys restrict an agent to a project or a function: a release-notes agent never needs write access to the CRM pipeline.

Because the server speaks the same state model as the rest of Atur, an agent-driven triage does not create a shadow workflow. A ticket the agent sets to “waiting on customer” is exactly the state the human teammate sees the next morning, with the SLA clock paused in the same place.

The safety habit

Agents make mistakes at the speed of token generation, so the rule is to review, not delegate. Every agent action is a diff that a human can approve, and scoped keys make a runaway agent’s reach small. Teams that start with read-plus-draft (summarize, draft replies, suggest states) before granting write access report far more confidence than teams that flip the switch on day one.

For the kanban discipline that gives those AI-suggested states something stable to act on, read how WIP limits keep a board moving, the review habit in AI ticket triage and agent-drafted release notes, and the workflow section of the site for how states, projects, and cycles fit together.