Trello MCP server vs Jira and Linear for AI coding agents

Since July 2026, Trello has an official MCP server. Atlassian’s server covers Jira, and Linear has had one for a while. So the obvious move for anyone running Claude Code, Codex or Cursor is to point the agent at the board you already use and let it pull work from there.

Sometimes that works. If you want an assistant that reads your board, files cards and tidies things up while you talk to it, any of these servers will do. If you want coding agents to take work off the board on their own, several at a time, you’ll run into problems quickly. This post covers how each server works, how to connect it to Claude Code, and where the gaps are.

What a board MCP server actually does

The Model Context Protocol lets an AI client call tools on a remote server. A board’s MCP server wraps its API as tools like “search cards”, “create card” and “move card”, and the client decides when to call them.

Two things follow from that:

• The agent only acts while a session is running. Nothing happens on the board unless an agent, mid conversation, decides to call a tool.
• The board only knows what the agent tells it. If the agent forgets to move the card, the card doesn’t move.

Keep both in mind as you read the rest.

Trello’s official MCP server

Trello’s server lives at https://mcp.trello.com/v1 and uses OAuth 2.0, so you sign in with your Trello account and pick which workspace the assistant can see. It works on every Trello plan. The one exception is creating focus time in Planner, which needs Premium or Enterprise.

What it can do: read boards, lists, cards, checklists, comments, custom fields, attachments and members; create and update cards; move cards between lists; manage checklists; add comments; assign members; attach URLs; mark cards complete; archive cards and lists; and search by keyword. It also reaches Trello’s Inbox and Planner.

What it can’t do: permanently delete anything. Atlassian’s documentation is explicit that the assistant “can archive cards and lists, but cannot permanently delete them”. Uploading files from your machine isn’t supported yet either.

To connect it to Claude Code:

claude mcp add --transport http trello https://mcp.trello.com/v1

Then run /mcp inside a Claude Code session and complete the sign-in in your browser. In the Claude desktop app or on claude.ai, add it as a custom connector with the same URL.

Before the official server there were community Trello MCP servers built on Trello’s REST API with an API key and token. They still work, but OAuth is the better default: you can see and revoke what the assistant was granted without regenerating keys.

Jira: Atlassian’s MCP server

Atlassian’s own server covers Jira, Confluence, Jira Service Management, Bitbucket Cloud, Compass, Loom and more. It does not cover Trello, which has its own server above.

claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp

Run /mcp to authenticate with OAuth 2.1. For headless or service-account use, an admin can enable API token authentication instead. If you set this up last year, check your config: the old SSE endpoint (https://mcp.atlassian.com/v1/sse) stopped being supported after 30 June 2026.

Jira’s strength here is its data model. Agents can read custom fields, sprints, links between issues and Confluence pages, and a well-kept Jira project gives an agent plenty of context. The downside is the same data model: a workflow with ten statuses and required fields is a lot for an agent to handle correctly on every transition.

Linear’s MCP server

claude mcp add --transport http linear-server https://mcp.linear.app/mcp

Linear’s server finds, creates and updates issues, projects and comments, and handles milestones and issue relations. It supports OAuth 2.1, or a Linear API key as a bearer token. There’s also a read-only endpoint at https://mcp.linear.app/mcp/readonly, which is handy for an agent you only want planning or reporting.

Linear has also gone a step further. Since June 2026, coding sessions let you hand an issue to an agent that runs in a managed sandbox inside Linear, on the Basic, Business and Enterprise plans, paid for with AI credits. That’s a different model from the rest of this post: the agent runs in Linear’s cloud rather than in your terminal, against your repository.

Side by side

All three are good at the job they were designed for: letting an assistant work with a board that people run.

What breaks when coding agents work from a people board

The trouble starts when the agent is no longer something you talk to but a worker that pulls tickets. Here’s what tends to go wrong, whichever of the three you use.

1. Nothing stops two agents taking the same card

Start two Claude Code sessions with “take the top card in Ready” and there’s a good chance both read the same card before either one moves it. Neither board’s MCP server offers an atomic “claim this card, and fail if someone else already has it”. You can ask agents to move the card first, but that’s a convention, and read-then-move leaves a gap where both agents get the same card.

2. Every agent is you

OAuth signs the assistant in as you. Three agents in three terminals all act as the same Trello or Linear user, so the activity log says you moved every card. You can’t see which session is on what, and you can’t cut off one misbehaving agent without cutting off all of them.

3. There’s no “I’m stuck” state

When an agent hits a decision it can’t make, it asks in the terminal. The card doesn’t change. Unless you’re watching that terminal, you find out an hour later that the agent has been waiting since it started. You can tell it to comment and move the card to a “Waiting on me” list, but that only happens if the model remembers, and nothing notifies you when it does.

4. Status is whatever the agent remembers to do

Moving a card is a separate tool call the model has to choose to make. Trello’s Butler automation, Jira workflows and Linear’s automations react to changes on the board, but none of them know that an agent started, finished or went quiet 40 minutes ago. If the agent crashes halfway through, the card stays in Doing until you notice.

5. Cards are written for people

“Fix the login bug” is a fine card for a teammate who was in the standup. For an agent it’s a guess. Boards built for people don’t push you to write down what “done” means, so agents either ask a lot or guess.

Making Trello, Jira or Linear work anyway

If you’re staying on your current board, you can close some of these gaps with conventions. Put something like this in your CLAUDE.md or AGENTS.md:

## Working from the board
- Only take cards from the "Agent: Ready" list. Never take a card that has a member on it.
- Before you change any code, add yourself as a member and move the card to "Agent: Doing".
- If you need a decision from me, comment the question on the card, move it to
  "Waiting on me", then stop.
- When done, comment a one-line summary with the PR link and move the card to "Review".

Then run one agent per list (or per label) so no two agents read the same queue. It isn’t airtight, but it works well for one or two agents you’re supervising.

When a coding agent needs a board built for agents

Once you’re running several agents, or leaving them unattended, you need the board to enforce these rules instead of relying on the model to follow them. That’s why we built Yokka. The difference is in the tools agents get:

• claim_card is a lock. A card is held by one agent at a time, and the claim is refused if another active agent holds it. If two agents ask for the next card at once, both may be shown the same one, but only the first claim wins: the second is refused and picks another.
• Every connection is its own agent. Each session shows on the board named after its client (Claude Code, Codex, Cursor), and each can have its own token: read-only or read-and-write, one project or all, with an expiry. Revoke one and only that agent stops. See Agent access.
• request_input is a state. When an agent asks a question, the card moves to Needs you and is flagged for you. You answer on the card and the agent carries on.
• The board reacts to what agents do. Flow rules move the card when an agent claims, asks, completes or releases it, in the same transaction as the call. An agent that goes silent for 30 minutes gets its card flagged.
• Cards have a brief. What, why and done-when, plus an optional checklist the agent treats as acceptance criteria.
• A card can start its own agent. A board MCP server waits for an agent to show up. With Yokka’s runner, a small open-source companion on the computer that holds your code, you press Start on a card (from the board or your phone) and it starts Claude Code or Codex there to work it, in place or in its own git worktree.

Connecting takes one command:

claude mcp add --transport http --scope user yokka https://api.yokka.ai/mcp/c/<token>

You don’t have to move everything. Plenty of teams keep the roadmap in Jira or Linear and put only the work agents pick up on a board built for agents. The quickstart connects your first agent in about five minutes, and the agent loop shows every call a session makes. If you’re running several Claude Code sessions at once, this guide covers worktrees, agent view and agent teams alongside the board.

FAQ

Does Trello have an official MCP server?

Yes. Trello launched an official MCP server in July 2026 at https://mcp.trello.com/v1. It uses OAuth 2.0 and works with Claude, Claude Code and other MCP clients.

Is the Trello MCP server free?

It works on every Trello plan, including Free. Creating focus time through Planner needs Premium or Enterprise.

Can Claude delete my Trello cards?

No. The official server can archive cards and lists but can’t permanently delete them.

Can Claude Code use Jira?

Yes, through Atlassian’s MCP server at https://mcp.atlassian.com/v2/mcp. Add it with claude mcp add and authenticate with /mcp.

Jira or Linear MCP for coding agents?

Linear’s server is simpler to work with and has a read-only endpoint, and Linear can also run coding agents itself. Jira gives agents richer context if your projects are well kept. Neither has an atomic claim, so both need conventions once more than one agent works from the same queue.

How do I stop two AI agents taking the same ticket?

You need a claim that fails if someone else already holds the ticket. On a people board, the nearest you’ll get is one agent per list or label. On an agent board like Yokka, claim_card does it for you.

Can a Trello card start a coding agent?

Not through the MCP server: it only responds when an agent calls it. You’d need your own automation to launch an agent when a card moves. On Yokka, pressing Start on a card has the runner start Claude Code or Codex on your machine.