/openapi-to-mcp

name: openapi-to-mcp
description: Build and deploy an MCP server from an OpenAPI / Swagger spec using the mcp-use TypeScript SDK. Use this skill whenever the user wants to "turn this OpenAPI spec into an MCP server", "make this API usable from Claude/ChatGPT", "wrap this Swagger doc as MCP tools", "expose this REST API to an LLM", "generate MCP tools from a spec", or pastes/attaches an `openapi.yaml`, `openapi.json`, or `swagger.json` and asks for a Claude-compatible version. Trigger even if the user doesn't say "MCP" — if they describe an existing HTTP API (REST endpoints, an internal service, a third-party API they have a key for) and want an LLM to call it, this is the right skill. Covers spec ingestion (file path, URL, or pasted), operation-to-tool mapping, auth wiring (apiKey, bearer, basic, OAuth bearer), scaffolding with `create-mcp-use-app`, tool generation with proper zod schemas, live testing in the mcp-use inspector, and deploying to Manufact / mcp-use cloud.

Build an MCP server from an OpenAPI spec

Turn an existing REST API — described by an OpenAPI 3.x or Swagger 2.0 document — into an MCP server. Each operation in the spec becomes one MCP tool the LLM can call. The server runs locally for testing and ships to Manufact / mcp-use cloud with one command.

This skill is the end-to-end recipe: scope → ingest spec → map operations → scaffold → generate tools → wire auth → test → deploy.

Core philosophy: the spec is the contract

The OpenAPI document is the source of truth. Tool names, descriptions, parameter shapes, and auth requirements all come from the spec — they should not be invented. This matters because:

• **The LLM trusts descriptions.** If the spec says `summary: "Get current weather for a city"`, that's exactly what the LLM will read when deciding whether to call the tool. Hand-rolled summaries drift; spec-derived summaries stay in sync if the API changes.
• **Zod schemas mirror OpenAPI schemas.** Every parameter — path, query, body — becomes a field in one zod object. Required/optional, enums, min/max, and descriptions all carry over. The LLM uses the schema to figure out what to ask the user for.
• **Auth lives outside the spec.** OpenAPI declares the auth scheme but never the secret. Secrets come from env vars; the spec tells you which env vars to require.

When in doubt, prefer mechanical fidelity to the spec over creativity. The LLM is doing the creative part — talking to the user — and only needs a faithful, well-typed handle on the API.

Process

1. Scope the request (use AskUserQuestion)

Before writing code, lock five things via the `AskUserQuestion` tool. All five are about the API and what to build — deployment is a separate question we ask later in step 10, when the user can actually evaluate it against a working server.

• **Spec source**: a file path in the workspace, a URL (e.g., `https://api.example.com/openapi.json`), or pasted into chat. If pasted, save it to `openapi.yaml` or `openapi.json` first.
• **Server base URL**: take it from `servers[0].url` in the spec if present; otherwise ask. Multiple `servers` entries are common (prod / staging) — confirm which one.
• **Auth scheme**: read `components.securitySchemes`. If multiple, ask which to use. If the API needs an API key or token, ask which env var should hold it (`API_KEY`, `OPENAI_API_KEY`, etc.). Don't ask for the secret itself — never put it in the conversation or commit it.
• **Operation filter**: large specs (Stripe, GitHub) have hundreds of endpoints. Ask whether to expose all operations, a tag (`pets`, `users`), or a hand-picked list. Default to "all" for specs under ~30 operations; ask above that. See `references/mapping-rules.md` for filtering patterns.
• **Widgets**: ask whether any operations should render a widget in the chat (a React component shown inline next to the LLM's reply), or whether this is a pure tools-only server. The default for an OpenAPI wrapper is **tools-only** — the LLM reads JSON and talks. Pick widgets only when the user wants a richer UI for specific responses (a map for a geocoding endpoint, a chart for a metrics endpoint, a card list for a search result). **This answer drives the scaffold template in step 3**: tools-only → `--template blank`, any widgets → `--template mcp-apps` (which ships the `resources/` widget infrastructure pre-wired). If the user wants widgets on most operations, the `build-mcp-app` skill is usually a better fit than this one — flag that and confirm before proceeding.

Don't skip this step. Generating 200 tools the user doesn't need pollutes the LLM's tool list and slows it down.

2. Acquire and dereference the spec

Get the spec into a single dereferenced JSON object on disk. Dereferencing inlines `$ref`s so downstream code never has to chase pointers.

# In the scaffolded project root
npm install @apidevtools/swagger-parser

// scripts/load-spec.ts (run once, manually or as a build step)
import SwaggerParser from "@apidevtools/swagger-parser";
import { writeFileSync } from "node:fs";

const spec = await SwaggerParser.dereference("./openapi.yaml");
writeFileSync("./openapi.dereferenced.json", JSON.stringify(spec, null, 2));

For URL specs, swagger-parser accepts the URL directly. For pasted YAML, write it to `openapi.yaml` first, then dereference. If the spec is Swagger 2.0, run it through `swagger2openapi` first (`npx swagger2openapi --outfile openapi.yaml swagger.yaml`).

Sanity-check the dereferenced file: open it, search for `"$ref"` — there should be none. If there are, the spec has circular refs and swagger-parser keeps them as-is; treat those refs as opaque object types in zod.

3. Scaffold with `create-mcp-use-app`

Pick the template based on the widget answer from step 1:

# Tools-only (the default for an OpenAPI wrapper)
npx create-mcp-use-app@latest <project-name> --template blank

# Any widgets at all
npx create-mcp-use-app@latest <project-name> --template mcp-apps