MCP Gateway

Running. This is an API, not a web application — there is nothing here to click.

It speaks the Model Context Protocol over HTTP at POST /mcp. Point an MCP client at it; a browser will only ever see this page.

Connect

{
  "mcpServers": {
    "mcp": { "url": "https://mcp.ceruleanai.com.au/mcp" }
  }
}

Your client discovers everything else on its own. The first unauthenticated call returns 401 with a WWW-Authenticate header naming the resource metadata, which names the authorisation server. Nothing needs configuring in advance.

Signing in

A person is sent to the identity provider in their system browser — authorization code with PKCE on a loopback redirect. The client never handles a password.

An agent with no human present uses client credentials and never opens a browser. The two are separate credential sets by design: a headless job cannot borrow a person's identity.

What you can call

Tools are grouped, and each group is authorised separately. tools/list returns only the groups your identity holds a role for, so this page deliberately does not enumerate them — ask the gateway with a token and it will tell you.

Endpoints

POST /mcpthe protocol endpoint — tools only
GET /healthzliveness and the protocol revision served
GET /.well-known/oauth-protected-resource RFC 9728 discovery

Getting a tool onto it

Two ways in. Register an existing API — choose the operations that should be agent-callable, write their descriptions, and the gateway publishes them as tools; you write no code. Or build a server from the template and deploy it into your own environment, where it stays yours to run.

Either way the contract is reviewed and the conformance gate has to pass before anything is reachable. Tool descriptions get the most scrutiny: they are read by a model choosing which tool to call, so they are written, not generated.