Connectors

A session changes the code in a workspace. The work around the code lives somewhere else: the ticket in Linear, the error in Sentry, the deploy in Vercel, the screens to follow in Mobbin. Connectors bring those services to the agent.

Where the work is#

A task touches several services before and after the diff.

Without connectors, the person copies each piece into the conversation and the result back out. The agent works from a summary, and every step waits on a person. With connectors, the agent calls those services itself, so a factory can run the whole task, not only the diff.

Added once#

A connector is an MCP server. It is added once on the Connectors page, and every new session's agent gets its tools beside its own. A session already running keeps the tools it started with.

The catalog holds hosted servers that sign in through the browser in one click. A server that isn't in it is added by its URL, or as a command the Mac runs.

Tokens stay in the factory#

An agent never holds a service's token. Each remote connector runs as a small local program that forwards the agent's calls to the factory, and the factory adds the sign-in on the way out.

The agent's transcript and environment never contain a token, and a token that expires mid-task is renewed without stopping the session.