How an agent signs in to Cairn and proposes changes

An agent that works with your roadmap needs three things. It needs a way to sign in that does not involve pasting a secret. It needs answers it can read whole. And it needs a way to try a change before anything is saved. Here is how the Cairn MCP server does each one.

Sign in without copying a token

Cairn is its own authorization server for MCP. A client that follows the MCP authorization flow connects with one command that needs only a name and the address of your Cairn. The connection guide below prints the commands with the right address.

Such a client finds out how to sign in, registers itself and opens your browser. You sign in with Google, then Cairn shows what the agent asks for, in plain sentences, and you allow or deny. Scripts and CI jobs that cannot open a browser can use a token instead.

We test this flow with the official MCP SDK client. We have not tested every client yet, so if yours does not connect, tell us.

The agent acts as you, with your role as it is now. If an administrator lowers your role, the next request of the agent already has the lower role. If you are removed, the agent stops. An access token lasts one hour, and a refresh token lasts 30 days from its last use. The settings page lists every connection, and revoking one ends it at the next request. The change feed names the member and the client behind each change.

Answers an agent can read whole

An agent cannot use an answer that it has to save to a file. So every read of the server answers within about 20,000 characters, even on a roadmap of 200 projects. Long lists come in pages with a cursor. When a result is cut, it says so and names the call that returns the rest.

Try a change before it saves

The tools that write projects and programmes accept a dryRun flag. With it, the agent gets the difference that the change would make and nothing is saved. It reads the difference, then sends the same call again without the flag. A bulk change of more than 20 records returns a plan until it is confirmed.

What an agent can and cannot do

At the time of writing the server has 23 tools. An agent can read projects, programmes, changes and the summary of the roadmap, search it, create and update projects and programmes, archive them, and manage share links. It never does more than the person who connected it: an agent of a viewer can only read. It cannot delete a project, because that action has no tool and is reserved for administrators in the app.

Connect your own

The connection guide has the commands and the steps for a token, and the list of tools.