MCP servers from your OpenAPI spec

1. 1

   Give Statio your spec

   Paste the URL of an OpenAPI 3 spec. Statio fetches it, counts the endpoints and shows the tools it will generate, along with the base URL those tools will call, which you can change.
2. 2

   Pick a generation mode

   One tool per endpoint for small APIs, or three generic tools for large ones. The modes are described below.
3. 3

   Ship a version

   Statio generates the server, builds the image, deploys it privately on Google Cloud Run, discovers its tools and moves traffic to it. You can follow each stage as it runs.
4. 4

   Grant it, tool by tool

   Every generated tool becomes something you can grant or deny per member and per agent, like any other server behind the gateway.

Most MCP servers register every tool up front. When a model connects, it receives the full catalog: every tool name, description and input schema, placed in its context window before the user has asked anything. A server with hundreds of tools can spend tens of thousands of tokens that way.

That cost is why some teams give up on MCP for raw CLI wrappers, and lose access control, credential handling and the record of every call along the way. Generation modes are how Statio keeps MCP cheap for large APIs.

• Per Endpoint

  One tool per endpoint

  Every operation in the spec becomes its own tool, named from its operationId, with its path and query parameters typed from the spec. Best for small, focused APIs where you want the model to see the whole surface up front.

  Filter by tag, path, method or operation to keep only the endpoints you want.

• Generic

  Three tools, whatever the size of the API

  The server registers exactly three tools, whether the API has 20 endpoints or 1,000. The model browses, reads the schema it needs, then calls. Everything else stays on the server in an endpoint map and is only sent when asked for.

  The list of tools the model reads on connect stays the same size as the API grows: an estimated 800 tokens.

• Curated

  Hand-picked tools, with a fallback

  Name the operations that matter as direct tools, with your own names and descriptions. A fallback tool still reaches every other endpoint, and an explorer and schema tool help the model find them. Curated mode is configured in mcp-gen.yaml with the CLI today.

  generation: mode: curated curated: tools: - operationId: createCustomer - operationId: getInvoice name: stripe_invoice - operationId: createPaymentIntent

An operation can be named by its operationId, by method and path, or by a concrete URL path.

• What the model usually has to know

  path: { petId: "123" } query: { status: "available" } headers: { "X-API-Key": "sk-..." }

• What a generated server asks for

  operation: "getPetById" params: { petId: "123", status: "available" } // Routed to path and query on the server. // Credentials are added outside the model.

• TypeScript on the official MCP SDK

  Generated servers are TypeScript, run on Bun, and use the official Model Context Protocol SDK over Streamable HTTP, with per-session state, idle sessions closed after 30 minutes and a cap on open sessions.

• Retries that never repeat a POST

  Calls time out, and failed calls are retried with exponential backoff and jitter on 5xx, 429 and 408. POST and PATCH are never retried, so a flaky upstream cannot charge a customer twice.

• Real-world specs

  References are resolved and inlined, including recursive and mutually recursive schemas. Operations without an operationId get a stable name from their method and path.

• Credentials stay out of the model

  The model never sees or supplies a credential. In Generic mode, a header the model passes can never override the authentication header.

• A small, non-root container

  Each server ships with a Dockerfile on a patched Alpine base that runs as a non-root user, plus a health endpoint.

• Versions, not redeploys

  Ship a new version and Statio fetches the spec again. If the spec, the generator and your settings are unchanged it reuses the last image; if anything changed it rebuilds and tells you why.

• Preview, promote, roll back

  Each version can run on its own preview URL before you promote it. Roll back to an earlier version, cancel a build in progress or rebuild one, from the server’s page.

• Private from the first deploy

  Generated servers are never public. Only the gateway can call them, and their outbound traffic goes through the egress proxy and that server’s allowlist.

• Scanned on the way in

  When the server’s tools are discovered, their names and descriptions are scanned for prompt injection and anything suspicious is flagged for an admin.

The config sets the spec (a URL or a local file), the server’s base URL and authentication, the generation mode and filters by tag, path, method and operation. Unknown keys are rejected, so a typo fails loudly instead of being ignored. Sync regenerates every file from the spec, so keep your changes in the config.