MCP Server for Project Management: a 10-Point Checklist, Scored

A good MCP server for project management lets an AI client read and change your board without a key you have to babysit, lets you cut off one client without touching the rest, and gives agents a safe way to claim work, ask you questions and prove what they did. Below are ten checks that cover this, why each one matters, and a scored table for Weft, Hypertask, Linear, Trello and Notion.

We make Weft. Facts about other products come from their official pages, checked on 27 October 2026. Where a page does not say, the table says "Not stated".

Why score a project management MCP server at all?

The Model Context Protocol solves the plumbing: a client asks a server which tools exist and calls them with structured arguments. It says nothing about how long a token lives, whether two agents can grab the same task, or what happens when an agent wants to delete something. Those are product decisions, and they differ a lot between servers.

Hypertask's guide to project management MCP servers has a sensible evaluation list: tool coverage, schema quality, auth and scope, audit history, destructive actions, transport. It just never fills it in. This post takes a similar list, sharpens it into ten checks you can answer yes or no, and fills in the answers, including ours.

What are the ten checks?

1. How do you sign in, and how long does the token last?

A key you copy into a config file is a password sitting on disk, and when it expires your agent fails quietly in the middle of a run. OAuth with refresh tokens means nothing to copy and no calendar reminder to renew. Ask the lifetime of both the access token and the connection, not just the method.

2. Can you revoke one client without killing your account?

You will connect several clients: a chat app, a coding agent, maybe a phone. When one misbehaves or a laptop is lost, you want to cut that one connection and leave the others running. If the only lever is rotating a shared key, every client breaks at once.

3. Can each connection get its own permission level?

A client you use to ask "what is on my board?" does not need write access. A coding agent needs write but rarely needs bulk archive. Per-connection levels such as read, write and full let you give each client the least it needs, which is the single best defence against a confused or manipulated agent.

4. Are destructive actions marked and limited?

MCP tools can carry hints that say whether a tool is read-only or destructive, and clients can use them to ask before acting. Check that the server sets them honestly, and check what "delete" really does: a trash bin you can restore from is very different from a permanent delete.

5. Can an agent claim a task so two agents don't collide?

Moving a card to Doing is a visual signal, not a lock. Two agents reading the board a second apart will both see the task free and both start. A claim with a lease, decided by the server, means the second agent is told no, and a lease that lapses when an agent dies frees the task for someone else.

6. Can the agent ask you a structured question?

Agents hit forks only a person can decide. Without a tool for it, the agent either guesses or stops and the question dies in a chat log. A structured question with options, delivered somewhere you will actually see it, keeps the decision attached to the task.

7. Does completion carry evidence?

"Done" from an agent is a claim. A completion that records links, file hashes or other artifacts, plus what was not done, lets you check the work without re-deriving it. The board should also show which client did the work.

8. Which clients can actually reach it?

A server is only as useful as the clients you already use. Check the real route for each: a published app in ChatGPT, a custom connector or a directory listing in Claude, a one-line setup in Claude Code, Cursor and Codex. Also check whether the agent's question can reach you, the person, on your phone.

9. What does the free plan allow over MCP?

Agents make many small calls. A free plan that counts boards, tasks or calls can run out mid-week. Find the exact limit, and whether MCP is included on the free plan at all, before you build a habit on it.

10. Is the transport current?

The MCP spec moves. The current remote transport is Streamable HTTP, and older SSE endpoints are being deprecated by several vendors. A server that names the spec revisions it supports is easier to trust when your client updates.

How do Weft, Hypertask, Linear, Trello and Notion score?

Each cell uses only what the vendor's own page or live server says. "Not stated" does not mean the feature is missing; it means we could not confirm it from an official source today.

Some honest notes on reading this. Linear, Trello and Notion are broad products whose MCP servers are one surface among many, and their MCP docs do not describe agent-workflow tools such as leases or structured questions, so many agent-specific rows are blank. Trello's refusal to expose permanent deletes is a good safety choice. Linear's separate read-only endpoint is a clean way to do check 3. If your team already lives in one of them, the official server is usually the right start; see Weft vs Linear, Weft vs Trello and Weft vs Notion.

Hypertask builds lease, evidence and escalation into its product, but its MCP docs say those live in the REST API, not in MCP, by design. That works well if your agents run scripts. It means a chat client connected over MCP alone does not get them. Hypertask also has things Weft does not: a CLI for scripts and CI, a REST API, agents as service accounts with their own identity, built-in AI features, time tracking and a keyboard-first interface. Our longer comparison is at Hypertask alternative.

What does Weft not score well on?

Weft's MCP server has 32 tools, and listing them is not cheap: we measured the tools/list response at about 17.7k tokens. Hypertask's tools reference groups its tools by action, so 21 tools cover more operations, each with an action field. That is a real tradeoff. Grouped tools take less room in the model's context. Separate tools let every tool carry its own read-only or destructive hint, which is what makes a read connection mean exactly 11 read-only tools and a write connection mean everything but the 3 destructive ones. We chose the second; if your client is tight on context, the first may suit you better.

Other gaps, plainly. Weft has no CLI, no public REST API and no webhooks, so it cannot be driven from a CI job without an AI client. There is no GitHub or Slack integration. Personal access tokens have no scopes: one token reaches all 32 tools across the workspace, so prefer OAuth where your client supports it, because per-connection levels apply to OAuth connections only.

How should you guard against prompt injection?

An agent that reads task text, comments or linked pages can be steered by instructions hidden inside them. MCP does not fix that; your setup does. Three habits help.

• Explore on a read-only connection. When you only want summaries, standups or search, connect that client at the read level. In Weft that leaves only the 11 read-only tools, and a refused call does not count against your quota.
• Review destructive calls. Keep your client's confirmation on for tools marked destructive instead of approving them for the whole session. In Weft those are trash_task, archive_done_tasks and complete_sprint, and trashed tasks stay restorable for 30 days.
• Treat board content as data. Instructions in a task description are not instructions from you. Put the rules your agent follows in its system prompt or project rules, where task text cannot overrule them.

The claim loop helps too. When an agent reaches an irreversible step such as publishing or touching production, it can call request_input and wait for your answer instead of acting on text it found in a card.

When don't you need an MCP server?

MCP earns its overhead when a model has to choose among actions on a board that keeps changing. Many jobs are not like that.

• One event always produces one known update: a script against the vendor's API is simpler and cheaper than a model.
• You need a periodic snapshot: an export does the job. Weft offers JSON and Markdown export and a private calendar feed.
• One person, one short session: a Markdown checklist in the repo is fine until work outlives the chat.
• Your pipeline runs in CI with no AI client: pick a tool with a CLI or REST API. Weft has neither today.

How do you test a server before you commit?

1. Connect one client at the read level and ask it to summarise the board. Nothing should change.
2. Switch to write on a test project and have it create one task and move it. Check the board shows which client did it.
3. Start the same task from two clients. The second should be refused, not silently duplicate the work.
4. Have the agent ask you a question with options, and check where it reaches you: the app, your phone, your email.
5. Finish the task with a link as evidence, then disconnect the client and confirm its next call fails.

In Weft, step 1 is one command in Claude Code, one click in Cursor, a config line plus codex mcp login weft in Codex, or adding the app in ChatGPT. The full field of MCP task managers is at MCP task managers compared, and the claim loop is explained in the claim loop.

Isn't tool count a sign of quality?

No. A long tool list costs context on every session; Weft's 32 tools measure about 17.7k tokens. What matters is whether the tools you need exist, carry honest read-only and destructive hints, and fail with clear errors.

Is OAuth really safer than an API key?

It removes the copied secret and lets you revoke one client on its own. Weft issues refresh tokens, so the client renews the sign-in itself. For clients without OAuth, Weft also offers personal access tokens, but those have no scopes.

Can I give ChatGPT read-only access to my Weft board?

Yes. Choose the read level on the consent screen, or change it later in Settings under Connected AI clients. That connection then sees only the 11 read-only tools.

What happens if an agent crashes while holding a task?

In Weft the lease lapses after 20 minutes without a heartbeat and the task goes back to unassigned, so another agent or person can pick it up.

Do I need a paid plan to try this?

No. Weft's free plan includes 50 active tasks and 200 MCP calls a month, and Pro is $7 per seat a month or $60 a year. See pricing.

We already use Linear or Trello. Should we switch?

Not just for MCP. Their official servers work on the data your team already has. Consider Weft if you want claims, questions and evidence inside MCP itself; import from Linear CSV or Trello JSON is free on every plan. See migrate from Linear or migrate from Trello.