Integrations

How Beacon works with the tools around it. The pages come in two groups: the agent clients that connect to Beacon, and the events and notifications that pass between Beacon and other systems when something breaks.

Agent clients

An agent reaches Beacon over MCP or the REST API as one of a person's agents, and never does more than that person's role allows. An interactive client signs in through the browser with MCP's OAuth; a program that runs alone uses a token.

1. Claude Code: sign in through the browser, and reach machines with your own ssh, scp and rsync.
2. Codex: the same sign-in, and a Beacon without the internet.
3. Cursor: sign in through the browser, or send a token.
4. Other MCP clients: any client that speaks MCP over HTTP, and the two ways to sign in.
5. Headless agents and CI: an agent's token, beacon-oncall to run an agent when a hook fires, and tokens that renew themselves.
6. The on-call prompt: the prompt a woken agent starts from.

Events and notifications

What goes out from Beacon when an alert fires, and what comes in to explain it.

1. Agents woken by a webhook: a signed POST to a receiver of yours, and how to check it.
2. GitHub Actions: a hook that dispatches a workflow, and a workflow that runs Claude Code.
3. Deployments and Flux: change events from CI and Flux, so that the evidence says what changed just before an alert.
4. Set up the Feishu app: alerts as cards in a Feishu or Lark group, with buttons for the people on call.

When a hook fires and what the agent does next is in Put an agent on call; what a hook sends, field by field, is in the Webhooks reference.