Which PM Tool Lets Your AI Agent Actually Do Things?

You want your AI assistant to run your project board. Not summarise it, not describe it — actually create the card, move it, assign it, and leave the comment explaining why. Every project-management vendor now says it supports this. Most of the comparison pages you will find rank them by whether they have a Model Context Protocol server at all.

That question is settled. On 2026-07-29 we sent direct HTTP probes to the official MCP endpoints of eight competing project tools. All eight responded. All eight were live and OAuth-protected. None of them were missing.

This page settles the question that is actually open: once your assistant connects, what is it allowed to do? Everything below comes from that probe or from vendor documentation quoted directly. Where we could not verify something, we left it out rather than guess.

The short answer

Every major project tool shipped an official MCP server, so MCP support is no longer a differentiator. It is table stakes as of 2026-07-29.

The differentiator is permission depth. An MCP server can expose a read-only window onto your data, or it can expose a full write surface that lets an agent create, move, comment on and delete work items. Two servers can both be "live and OAuth-protected" and still differ enormously in what a connected assistant can accomplish without a human clicking anything.

Three specific things separate the servers we looked at: how granular the OAuth scope model is, whether one connection reaches one workspace or many, and whether destructive or communicative actions (delete, comment) are permitted at all. Those are the axes worth comparing.

Every major project tool now has an MCP server

All of the following endpoints were probed live on 2026-07-29 and returned an OAuth-protected MCP server. Beta status is noted where the vendor labels it as such.

• Zoobbe — https://api.zoobbe.com/mcp
• Trello — https://mcp.trello.com/v1 (launched 2026-07-22)
• Linear — https://mcp.linear.app/mcp
• Notion — https://mcp.notion.com/mcp
• Asana — https://mcp.asana.com/v2/mcp
• ClickUp — https://mcp.clickup.com/mcp (beta)
• Monday.com — https://mcp.monday.com/mcp (beta)
• Shortcut — https://mcp.shortcut.com/mcp
• Jira — https://mcp.atlassian.com/v1/mcp, the Atlassian Rovo server, covering Jira, Confluence, Bitbucket and Compass

One clarification that trips people up: Atlassian runs two separate servers. The Rovo endpoint covers Jira and its siblings. Trello has its own endpoint at mcp.trello.com and is not included in Rovo. Connecting to one does not give your assistant the other.

Asana's V1 beta endpoint at /sse is documented by the vendor as shutting down on 2026-08-05. We did not verify that shutdown by probe — we are repeating the vendor's own note. If you built against the beta path, check it.

Some comparison articles still list one or more of these tools as having no MCP support. Those pages are out of date. This landscape moved fast in mid-2026, and a page written even a few weeks earlier would have been correct at the time it was published.

What actually differs: what the agent can do

The meaningful differences sit in the authorization model, not the protocol. Every server here speaks MCP over HTTP and gates access with OAuth. What varies is how finely that access is sliced and what falls inside it.

Scope granularity

Scope counts varied more than anything else we measured. Trello declares 12 OAuth scopes — the most granular model of any server we probed, more granular than Zoobbe's 11. Linear declares 2. ClickUp declares 2. Notion, Asana and Shortcut each declare 1. Monday.com and Atlassian's Rovo server did not declare scopes_supported in their protected resource metadata at all.

Granularity matters because it is the only mechanism you have for saying "this agent may read cards but may not touch members" or "this integration gets attachments and nothing else." A single coarse scope means every connected assistant gets everything the token covers. If least-privilege access is a requirement in your organisation, Trello's model is the strongest one we found, and it is worth saying that plainly.

Multi-workspace reach

A connection that spans workspaces changes what agent workflows are possible. Zoobbe's server is multi-workspace: one authorised connection can reach across the workspaces the user has access to. Trello's is not. From Atlassian's own support documentation:

Each Trello MCP connection supports one workspace. Multi-workspace support is planned for a future release.

If you run one workspace, this is invisible. If you run a workspace per client, per product line or per department, it means one connection per workspace and no cross-workspace query.

Write depth: comments and deletes

Read access is cheap; write access is where servers diverge. Two write capabilities are worth checking specifically, because they are the ones agents reach for constantly.

Commenting is how an agent explains itself. An agent that moves a card but cannot say why leaves your team guessing. Trello lists comments — along with attachments, custom fields and activity history — under "What's coming" as of 2026-07-29. Zoobbe's agent can post comments today.

Deleting is the other one. Trello's documentation is explicit:

Destructive delete operations are not supported. Your assistant can archive cards and lists, but cannot permanently delete them.

Zoobbe's agent can delete. Which of those you want is a genuine judgement call, and we cover it below.

We are not publishing tool counts for any vendor except ourselves. Zoobbe's hosted server exposes 65 tools, verified in our own source, with the hosted server and the CLI at full parity — the same 65 in both. We did not verify tool counts for any competitor and will not estimate them. Counts are a weak proxy for capability anyway; a server with fewer, better-shaped tools can beat a server with more.

What this looks like in Zoobbe

Connect Zoobbe to Claude, ChatGPT, Cursor or any MCP client and the assistant works against your real boards through 65 tools. In practice that means you stop describing work to your assistant and start delegating it.

• “Move everything blocked on design review into next sprint and tell the owners why.” The agent reads the board, moves the cards, and posts a comment on each one — comments are a first-class tool, not a planned feature.
• “What did the team actually ship last week?” It reads board activity and card history across every workspace you granted, not one at a time.
• “Set up the launch board from our template and assign the usual owners.” Boards, lists, labels, members and templates are all addressable.
• “Start a timer on the card I am working on and add it to My Day.” Time tracking and the personal daily list are exposed to the agent too.

The same 65 tools are available through the hosted server and the Zoobbe CLI, so whatever you script locally behaves identically to what your assistant does remotely. Authorisation is OAuth 2.0 with PKCE and dynamic client registration, so connecting a client is a consent screen, not an API key pasted into a config file.

Zoobbe and Trello, side by side

Trello is the closest comparison because it is the tool most Zoobbe users are coming from, and because Atlassian documented its MCP limits openly enough to compare honestly. Everything in this table reflects the state on 2026-07-29, seven days after Trello's server launched.

Zoobbe's server also supports PKCE with S256, dynamic client registration and RFC 9728 discovery, so a compliant MCP client can find and register against it without manual credential setup.

Why Trello's limits are not a bug

Atlassian blocked destructive deletes on purpose, and the reasoning is sound. An agent that misreads an instruction and archives twenty cards is an annoyance you undo in a minute; the same agent with delete permission is data loss. Scoping a connection to one workspace follows the same logic — a prompt injection buried in a card description cannot walk into workspaces the connection was never granted.

We took the opposite trade deliberately. Zoobbe puts delete behind its own explicit scope, so you grant it only when you want it, and every action an agent takes lands in the same board activity trail as a human edit. You get the reach, and you keep the receipts. That is a difference in risk appetite, not a claim that Atlassian shipped something broken.

What this means if you are choosing today

Pick on permission depth, because MCP availability will not decide it for you. Three cases cover most readers.

Stay on Trello if you run a single workspace, want the agent's worst-case action to be reversible, and need fine-grained scopes to satisfy a security review. That combination is exactly what Trello's server was built for, and nothing here changes it. Read-and-organise workflows — triage, sorting, status queries, moving cards — work today.

Look harder if your agent workflows need to span workspaces, or if you need the agent to leave a comment explaining its own actions. Those two gaps are the ones most likely to stop an automation in the first week. Both are on Atlassian's published roadmap, so the question is whether you can wait.

Zoobbe fits when you want an agent operating with the same permissions a teammate would have: create and move cards across boards, set priorities and due dates, post threaded comments with @mentions, work across multiple workspaces on one connection, and clean up when something is genuinely wrong. The 65 tools on the hosted server are the same 65 in the CLI, so a workflow you prototype locally behaves identically when an assistant runs it.

Date-stamp all of this. Trello's server was seven days old when we probed it, and Atlassian has published comments, attachments, custom fields and activity history as coming. This is a snapshot taken on 2026-07-29, not a permanent verdict, and we will re-probe.

FAQ

Which project management tools have an MCP server?

All the major ones. As of 2026-07-29 we probed live, OAuth-protected official MCP servers for Zoobbe, Trello, Linear, Notion, Asana, ClickUp, Monday.com, Shortcut and Atlassian's Rovo server covering Jira. Any comparison claiming a major PM tool has no MCP server is out of date.

Can Claude or ChatGPT actually create and move cards, or only read them?

It depends entirely on the server's write surface. Zoobbe's server permits create, move, comment and delete. Trello's permits card and list changes including archiving, but not permanent deletion, and not comments as of 2026-07-29.

Does Trello's MCP server support multiple workspaces?

No, not as of 2026-07-29. Atlassian's support documentation states that each Trello MCP connection supports one workspace and that multi-workspace support is planned for a future release.

Why can't my AI assistant delete Trello cards?

Because Atlassian deliberately excluded destructive deletes. Assistants can archive cards and lists instead, which keeps every agent action reversible. It is a safety decision rather than an unfinished feature.

Is Jira's MCP server the same as Trello's?

No. Atlassian runs two separate servers: the Rovo server at mcp.atlassian.com/v1/mcp covers Jira, Confluence, Bitbucket and Compass, while Trello has its own endpoint at mcp.trello.com/v1. Connecting to one does not grant access to the other.

How many tools does Zoobbe's MCP server expose?

65, verified in our own source, with the hosted server and the CLI at full parity. We are not publishing tool counts for other vendors because we did not verify them.

Connect Zoobbe to Claude, ChatGPT or Cursor on the free plan and see what your assistant can actually do with a real board before you move anything. No card is required to try it. Start free at zoobbe.com.