Trello MCP: Two Servers, and Only One Reads Comments

Trello shipped an official MCP server on July 22, 2026. It is hosted at https://mcp.trello.com/v1, it authenticates with OAuth, and it works on every Trello plan including free. After a year of this search term pointing at community repos of wildly different quality, there is finally an obvious answer.

Install the official one. That is the right call for most people and this post is not going to talk you out of it.

But there is a gap in it that nobody on the results page mentions, and you will hit it inside your first ten minutes. Ask your agent what is overdue across your boards and it answers instantly. Ask it why the card labeled “Blocked” has been sitting there for three weeks and it cannot tell you, because it cannot read a single comment on that card.

The capability table everyone quotes is the wrong table. The one worth reading is the list underneath it, the one headed “coming soon”.

Two Atlassian MCP servers, and only one reaches Trello

The Model Context Protocol, or MCP, is the open standard that lets an AI client call outside tools mid-conversation. If you want the ground-up version first, we wrote a plain English explainer on what MCP is.

Atlassian runs two of them and the naming does nobody any favors.

The Atlassian Rovo MCP server covers Jira, Confluence, Compass and Bitbucket. It is scoped to an Atlassian Cloud site. It is the one most people find first, because it is the one Atlassian markets to companies.

Trello MCP is a separate server on a separate endpoint, scoped to a single Trello workspace. Atlassian’s own comparison table in the repo says Trello data “may also be available” through Rovo in a future release, which is a polite way of confirming it is not available there today.

So if you already wired up Rovo for Jira and assumed your boards came along, they did not. That is a five-minute mistake that reads like a permissions bug.

What the official server actually does

Here is the shipped surface, taken from the official repo rather than from a directory listing.

Boards and lists: view and create boards, view labels, view and move lists.

Cards: view, create, update, move, archive, mark done, attach and detach labels.

Checklists: view, create and update checklists, add and update items.

Search: find cards and boards by keyword.

Planner: view calendar events, create focus-time blocks, link and unlink cards to events.

Inbox: view, create, update and archive Inbox cards.

Three details are worth having straight before you connect.

Every plan works. Trello’s own support doc is explicit that there is no paid requirement. The single exception is creating focus time in Planner, which needs Premium or Enterprise. Reading calendar events from a connected Google or Outlook calendar is available more widely.

One workspace per connection. Not one at a time. One, full stop, at launch. If your work board and your side project live in different workspaces, an agent sees whichever one you authorized and has no idea the other exists.

Nothing gets deleted. Not cards, not lists, not labels, not comments, not attachments. The tools to do it were never built. An over-eager agent can bury a list in the archive, and you get it back from the board menu in about four clicks.

That last one deserves more credit than it gets. Handing an agent delete permissions on a shared board and trusting the prompt is not a safety model, which is the argument we made at length in how to vet an MCP server before you connect it.

Agent skill first

One thing in the official repo is genuinely useful and almost nobody mentions it. It ships an agent skill called trello-use that teaches a client the ARI id format the tools expect, how to handle UTC dates, and when to use Inbox tools instead of board tools.

npx skills install atlassian/trello-mcp-server

Load it before your first real prompt. Without it you spend your first session watching an agent guess at id formats and apologize.

The list that matters is the one marked “coming soon”

Atlassian publishes its roadmap in the same README as the capability table, which is honest and rare. Read the two lists side by side and the shape of the product becomes clear.

Not shipped yet:

• Comments: view, add and edit comments on cards.
• Attachments: view, upload and download card attachments.
• Custom fields: view definitions, get and set values.
• Activity and history: action history for cards, lists and boards.
• Labels: create and update them, rather than just attaching existing ones.
• Members: view and manage board, workspace and card members.
• Boards: edit name and visibility, archive a board.
• Cards: copy a card.
• Workspaces: connect more than one.

There is a tenth item that is not on the roadmap at all. In the launch thread, an Atlassian staffer confirmed that supporting board guests in Trello MCP is not currently planned. If you are a guest on a client’s board rather than a workspace member, that board is invisible to your agent and will stay that way.

Why the missing comments cost more than they sound

This is the part I would want somebody to tell me before I connected anything.

A Trello card title is an index entry. “Fix login bug.” “Q3 pricing page.” “Renew the domain.” It tells you a thing exists. It tells you nothing about the thing.

Everything you would actually ask an agent about lives one layer down. Why this slipped. What the client said. Which of three options we picked and who objected. That material is in the comments, in the attachments somebody dragged on, and in the activity trail showing the card bouncing between two lists for a month.

So the honest description of the official server today is that it has hands and no memory. It can build you a board from a conversation, triage your Inbox, and block two hours tomorrow for the thing you keep avoiding. Those are real wins and they are worth the ten-minute setup on their own.

What it cannot do is answer a question about work that already happened. Ask it for a recap of the quarter and you get a list of card titles you moved to Done, which is a worse version of looking at the board.

That gap closes when comments ship. Atlassian said in the launch thread that they are actively working on it. Until then, plan around it rather than discovering it on a Friday afternoon.

The 57-tool community server

delorenj/mcp-server-trello is the one the directories rank, and for once the directories are pointing somewhere reasonable. It has 431 stars, an MIT license, and it was last pushed on August 4, 2026. It runs on Bun with npx and npm as fallbacks.

Fifty-seven tools, and the ones that matter here are exactly the ones the official server is missing.

Comments: add_comment, update_comment, delete_comment, get_card_comments.

Attachments: attach_image_to_card, attach_file_to_card, download_attachment. Image attachments come back as inline viewable data, so an agent can actually look at the screenshot somebody pasted onto a bug card.

Custom fields: get_board_custom_fields and update_card_custom_field. These need Trello Standard or higher, which is a Trello limit and not the server’s.

Activity: get_recent_activity, plus get_my_cards across boards.

It also does something the official server does not, which is handle Trello’s rate limits for you. Trello caps requests at 300 per 10 seconds per API key and 100 per 10 seconds per token. The server runs a token bucket and queues requests instead of failing them, which matters the moment an agent decides to read every card on a 200-card board.

The nicer touches are workflow-shaped rather than API-shaped. There is a dedicated acceptance-criteria extractor, checklist item search by description, and fractional-index list reordering. That is somebody who uses Trello for real, not somebody wrapping endpoints.

The auth trade is the actual decision

Capability is not what should decide this. Credentials are.

The official server uses OAuth 2.0. You pick a workspace on a consent screen, you grant read, write and search separately, and you revoke it whenever you want. Actions respect the permissions you already have in Trello, so the agent cannot do anything you could not do yourself. Organizations managed through admin.atlassian.com can control which MCP domains are allowed and what access level users may grant, using the same framework as the Rovo server.

The community server wants a Trello API key and token from trello.com/app-key, pasted into your client config as environment variables. That token does not expire. It is not workspace-scoped by default, it reaches everything your account reaches, and it sits in a plaintext config file that half the developers reading this have accidentally committed at some point.

There is a mitigation and you should use it. Set TRELLO_ALLOWED_WORKSPACES and the server will only expose the workspaces you list. Leave it empty and every workspace the token can see is fair game.

Then remember that this server does have delete tools. delete_comment and delete_checklist_item are real, and they do what they say. That is the price of the capability you came for.

This trade shows up on every board tool that has shipped an official server. ClickUp’s official server refuses API keys outright and takes OAuth or nothing, which solves the leaked-token problem and creates a different one for anything running unattended.

Which one you want

Rovo is in that table for one reason, which is that it does not reach Trello and people keep assuming it does.

Between the two that do, the rule is simple. Start with the official server. If a week goes by and you never once wanted the agent to read a comment, you are done and you never have to manage a token.

If you did want it, add the community server for that job specifically, scope it with the workspace allowlist, and keep the official one for anything that writes to a shared board.

Once you are connected, this is the prompt worth running first, because it tells you immediately which half of the product you are holding.

Copy this prompt

List every Trello card assigned to me that is overdue or due this week,
grouped by board. Then pick the one that has been sitting in the same
list longest and tell me everything you can find out about why it has
not moved. Say plainly which parts of the card you could not read.

The first half answers on either server. The second half is where the official one has to admit it is looking at a title and a due date.

What neither of them reaches

Both servers read what you filed. Neither touches what you saved.

The X thread about the pricing model you are arguing about on that card. The Reddit comment that solved the exact bug three months ago. The LinkedIn post from a competitor that made you create the board in the first place. The article that explained the approach you are now taking. None of that ever became a card. It got a tap on a bookmark icon in a different app each time and vanished.

That split is why the second-brain idea keeps disappointing people. The task layer and the notes layer get all the tooling, and the capture layer gets none, even though the capture layer is where most of what you found actually lives. We went into it properly in building a second brain from bookmarks.

It is the same shape of gap Miro MCP has with connectors and Todoist MCP has with project context. The server reads the object. The reasoning was somewhere else the whole time.

Where ContextBolt fits

Straight about the seam. ContextBolt does not integrate with Trello, and it is not trying to replace it.

It sits on the other side of the same gap. Trello holds what you planned. ContextBolt holds what you saved, the web pages you clipped and the posts and threads you kept from X, Reddit and LinkedIn. A Chrome extension captures them as you save them, tags and clusters them on arrival, and Pro exposes an MCP endpoint so Claude can search that pile by meaning in the same session it reads your board. The two run side by side in the same agent, one for what you planned and one for what you kept.

The design instinct is worth naming, because it runs in the opposite direction to this whole post. ContextBolt’s bookmark server keeps its tool list short. search_bookmarks, list_clusters, get_cluster_bookmarks, get_recent_bookmarks, export_bookmarks, save_bookmark. A server you leave connected in every session has to be cheap to leave connected, and 57 tools is not cheap.

Basic is free and capped at 150 bookmarks, with AI tagging, topic clusters and instant search on your device, which is enough to find out whether you would use it. Pro is $6 a month, removes the cap, and adds cloud sync and the MCP endpoint with its semantic search.

Worth doing?

Yes, and the setup is about five minutes because there is nothing to install.

Go in with the right expectation. Today’s official Trello MCP server is a very good pair of hands and a poor pair of eyes. Use it to capture, triage, restructure and schedule, and it will earn its place in your client config the first week.

Do not build a reporting habit on it yet. Anything that starts with “why” or “what happened” runs straight into the comment wall, and no prompt engineering gets around a tool that does not exist.

Check the roadmap table in the repo before you assume otherwise. It moved twice between June and August, and comments are the next thing on it.

And while the client config is open, put ContextBolt beside the Trello server for the research behind the cards. Basic is free up to 150 bookmarks.