mcp

mcp.gagarin.cloud — gagarin's API, as tools an agent can call.

npm install
npm run build
npm test                 # builds, then node --test over dist/
npm start                # the HTTP server on :8080
npm run stdio            # the same tools over a pipe

The Model Context Protocol is how an agent finds and calls a tool it was not built with. This is gagarin's, and it exists so that adding gagarin to a coding agent is a URL rather than an install — the one artefact that makes the platform installable rather than merely documented.

Two ways in, one implementation:

Signing in

Remote, with OAuth. Give the client the URL and nothing else:

https://mcp.gagarin.cloud/mcp

Claude, ChatGPT and Claude Code prompt for sign-in when they connect: the client opens a browser and the human signs in with GitHub or Google. A request with no credential answers 401 with a WWW-Authenticate header pointing at /.well-known/oauth-protected-resource/mcp, which names api.gagarin.cloud as the authorization server. This server decides nothing about a token itself; the engine does, on every call. GAGARIN_MCP_ORIGIN and GAGARIN_ISSUER override the two public names for a development setup; they are separate from GAGARIN_API, which in the cluster is the in-cluster Service and no address a client could sign in at.

Remote, with a credential in a header. For a client that cannot sign in over OAuth, or a machine that should not: a credential from gg login or gg creds create.

{
  "mcpServers": {
    "gagarin": {
      "type": "http",
      "url": "https://mcp.gagarin.cloud/mcp",
      "headers": { "Authorization": "Bearer <your gagarin credential>" }
    }
  }
}

Local, over stdio. npm run stdio from a clone runs the same tools over a pipe, reading the file gg login wrote, or GAGARIN_TOKEN if it is set. It is here to develop against, and is deliberately not published to npm: the point of this server is that adding gagarin to an agent is a URL and not an install, and a package on somebody's laptop is a second copy of src/tools.ts that goes stale the day a tool changes.

A credential that has expired or been revoked is caught at the door too, because a client signs in again only on an HTTP 401: every POST asks the engine /v1/whoami with the caller's token first, and the engine's 401 becomes a 401 with error="invalid_token". One extra in-cluster call per request, and no cache — a remembered token is a stored token, and this server stores none. Any other failure there (the engine unreachable, a 5xx) is not a sign-in problem, so the request goes on and each tool reports it with the engine's own code.

What is here

Project memory

Every project carries a memory: small, durable facts an agent saves about a codebase — a decision and why, a convention, a gotcha — so the next session does not rediscover them. It is a built-in of the project, like its registry, and these tools are the only way to reach it: the memory service has no address of its own, no gg command and no console page. The engine decides who may read (viewer) and who may write (editor), on every call, like everything else.

These answer the memory service's own text — a plain-text rendering packed to the budget — rather than the JSON around it. That is a narrower case of the rule below, not an exception to it: the rendering is still the engine's, and returning both would spend twice the tokens the service exists to save. A body without text comes back as JSON like any other answer.

It is a translator, and nothing else

It holds no credential of its own, has no database, and makes no decision the API does not make. Every tool is one call to api.gagarin.cloud carrying the caller's own bearer token, and every refusal is the engine's refusal passed through unedited.

That is the property everything else rests on: possessing this server grants nothing at all. It is what makes it safe to put a public endpoint in front of a single write gate, and nothing here may be changed in a way that weakens it.

Two consequences worth stating, because both look like omissions:

• Answers are the engine's JSON, not a summary of it. Every service, ledger line and connection already carries a sentence written by the engine, precisely so a terminal and a dashboard cannot describe the same row differently. A third renderer here would be a third opinion to keep in step.
• Errors are [code] message with a hint: line, which is what gg prints. One format, so an agent that has read the gagarin skill recognises what comes back here without being taught a second one. When the engine sends more than the envelope — a memory_duplicate carries the near-duplicates — that is appended after the hint rather than dropped.

What it deliberately cannot do

Four things need a machine, and offering them here would produce failures that read like platform faults:

• Build and push an image. gg ship — build, push and deploy fused — shells out to docker where the source is. deploy here runs an image that is already in gagarin's registry: a tag CI pushed, or a restatement. Getting one there is the CLI's job.
• Open a tunnel. gg connect binds a local port.
• Wait for a job to finish. run submits and returns a revision, like every other write here. gg run blocks until the run ends and exits with the script's own exit code, which is what a pipeline wants; over MCP you poll status for the phase and the code.
• Click an approval. That is a human with an inbox, by design.

There is also no list of resource types, sizes or scopes in this repository. The engine owns those and its refusals name them; a z.enum here would be a second list in a second repository, wrong on the day a type is added — a mistake this codebase has already made once, in gg, about this exact family of values.

Where it runs

In the Kubernetes cluster, next to the control plane, and not on Vercel beside the site and the console. Those two live outside Scaleway because their job is to still be there and say the platform is down. This one has nothing whatever to say when the API is unreachable; it is that API in another shape, so it belongs next to it and reaches it over the in-cluster Service rather than hairpinning out through the load balancer.

The pod runs as a ServiceAccount with no RBAC and no mounted token, on a read-only root filesystem. It needs none of them.

Every push to main builds the image and rolls it out — .github/workflows/deploy.yml. The credentials that can do that are secrets of the production environment, which only main may use, so a pull request never runs next to them. Merging to main is deploying.

Adding a tool

One server.registerTool call in src/tools.ts: a name, a description saying the rule a caller most needs, a zod shape, and one api.call. Then a test in src/tools.test.ts asserting the path, the method and the body — those are the only things it is possible to be quietly wrong about, and a deploy that PUTs to the wrong path answers 404 and reads like a missing service.

Do not add a tool for something the API does not do. This file has no business being the place a feature appears first.