Agents & MCP

Coding agents can work your Marca task board directly: read the open tasks, pull the full brief for one, claim it, report back in its thread, and hand it to a human for review. They connect through marca mcp, a Model Context Protocol server that ships with the Marca CLI.

An agent always acts as the person who logged in, with that person's role on the project. Everything it writes shows that person as the author, marked "via" the agent's label, so the board stays honest about who did what.

Set up

Authorize the machine. This opens the browser, where you pick the project, name the agent and choose how long its token lasts (7–90 days, 30 by default).
```
npx @marca/cli login --label "claude-code on my laptop"
```
Check who the agent will act as:
```
npx @marca/cli whoami
```
Add the server to your agent.

Claude Code
```
claude mcp add marca -- npx @marca/cli mcp
```
Cursor (.cursor/mcp.json or ~/.cursor/mcp.json)
```
{
  "mcpServers": {
    "marca": { "command": "npx", "args": ["@marca/cli", "mcp"] }
  }
}
```
Self-hosting? Add --endpoint=https://review.example.com to both commands, or set REVIEW_ENDPOINT.

The token is stored in ~/.marca/credentials, readable only by you. In a container or CI job, where that file doesn't exist, set MARCA_TOKEN and MARCA_PROJECT instead.

What the agent can do

Tool	What it does
`whoami`	Who the agent acts as, the role, whether it can edit tasks, and when the token expires.
`list_tasks`	Tasks on the board, open by default. Filter by status or to tasks assigned to you.
`list_items`	Open tasks plus page comments that haven't become tasks yet.
`get_task`	One task in full: route, selector, element text, attachments, comments, subtasks.
`get_task_brief`	The same brief the Export button produces, ready to work from.
`create_task`	File a new task.
`update_task`	Change status, priority, instruction or assignee; `assignToMe` claims it.
`bulk_update_status`	Move several tasks at once.
`add_task_comment`	Report back on a task's thread. Followers are notified.
`create_subtask`	Split a task into smaller steps.

Two read-only resources are also available: marca://project and marca://tasks/open.

A typical loop: list_tasks → get_task_brief → make the change → add_task_comment saying what changed → update_task to in_review so a person can check it in Marca.

Permissions

The agent can do exactly what its user can:

Owners, admins and members: everything in the table above.
Reviewers: read, create tasks and comment, but not change a task's status, edit it, or split it. Those tools return a clear refusal. Reviewers' logins also don't include an API key, so marca crawl isn't available to them.

Agent tokens are separate from the token the review widget uses on your site. The widget's token lives in the page, so it stays limited to commenting and creating tasks even for owners.

Revoking

Open Account → Agent access in the dashboard to see every agent token you've issued, when it was last used, and when it expires. Revoke cuts one off immediately: the agent's next call fails with a message telling it to run marca login again.