OpenAPI MCP

Turn any OpenAPI 3 spec into a working MCP server. One file, about 130 lines, no dependencies.

node openapi-mcp.mjs https://your-api.com/openapi.json

Every API endpoint becomes a tool that an AI client (Claude, Cursor, etc.) can call. There is no code generation and no per-endpoint files. The spec is read at startup, and the tools are built in memory.

Quick start

Requires Node 18 or newer.

git clone <this-repo>
cd openapi-mcp
node openapi-mcp.mjs sample/animechan.openapi.json

It prints openapi-mcp: 4 tools from "Animechan" -> ... and then waits silently for an MCP client. That is normal: it talks JSON-RPC over stdin/stdout and is not meant to be read by humans.

Use it with an AI client

Claude Desktop (local mode)

Add this to ~/Library/Application Support/Claude/claude_desktop_config.json, then restart the app:

{
  "mcpServers": {
    "animechan": {
      "command": "node",
      "args": ["/full/path/openapi-mcp.mjs", "/full/path/animechan.openapi.json"]
    }
  }
}

If the client can't find node, use the full path from which node.

Claude Code (remote mode)

Start the server, then register it:

MCP_AUTH_TOKEN=secret node openapi-mcp.mjs sample/animechan.openapi.json --http 8787

claude mcp add --transport http animechan http://localhost:8787/mcp \
  --header "Authorization: Bearer secret"

Then ask in plain English: "give me a random quote from Naruto".

Two modes

In HTTP mode, clients send JSON-RPC messages as POST http://host:<port>/mcp. Put --http <port> last.

Test it by hand:

curl -X POST localhost:8787/mcp -H 'Authorization: Bearer secret' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Options

node openapi-mcp.mjs <spec-url-or-file> [base-url] [--http <port>]

Big APIs can have hundreds of endpoints. Too many tools makes the model choose worse, so use TOOL_FILTER to keep the useful ones.

Authentication

The server reads securitySchemes from the spec and puts your secret in the right place. You only supply the secret as an environment variable:

Example for Claude Desktop:

{
  "mcpServers": {
    "my-api": {
      "command": "node",
      "args": ["/full/path/openapi-mcp.mjs", "https://my-api.com/openapi.json"],
      "env": { "API_TOKEN": "your-secret" }
    }
  }
}

Notes:

• If the spec declares a security scheme but no credentials are set, a warning is printed to stderr at startup.
• If the spec lists several schemes, the first one in the top-level security list is used (or the first one defined). Per-endpoint security overrides are not applied, so the secret is sent on every call.
• OAuth client credentials are sent as HTTP Basic auth to the token endpoint.
• The OAuth browser login (authorization-code flow) is not built in. Get a token or refresh token once with your provider's tools, then pass it in.
• If a provider rotates refresh tokens, the new one is kept in memory only, so it is lost on restart.

How it works

1. Load the spec and resolve $ref pointers.
2. Build tools: each operation becomes one MCP tool. operationId is the name, summary the description, and parameters plus request body the input schema.
3. Call: a tool call makes the real HTTP request and returns HTTP <status> followed by the response body (cut at 50,000 characters).
4. Serve the tools over JSON-RPC 2.0, via stdio or HTTP.

APIs without an OpenAPI spec

Many small public APIs only have HTML docs. For those, write a small spec by hand (see sample/animechan.openapi.json, 4 tools in about 35 lines), then run it the same way.

Limitations

• OpenAPI 3 only. Convert Swagger 2.0 specs first.
• JSON request bodies only. No file uploads.
• Auth: bearer, basic, API keys, and OAuth2 client-credentials or refresh-token. There is no interactive OAuth login, and one scheme is used per API.
• HTTP mode speaks plain HTTP (put HTTPS in front when deploying) and has no streaming, batch requests, or OAuth for clients connecting to the server (it only supports the MCP_AUTH_TOKEN bearer token). Some remote clients may require these.
• HTTP mode speaks plain HTTP (put HTTPS in front when deploying) and has no streaming, batch requests, or OAuth. Some remote clients may require these.
• If you deploy remotely, every user shares your server's IP, so upstream rate limits apply to all of them. Set MCP_AUTH_TOKEN, or anyone who finds the URL can use your API key.

Contributing

Issues and pull requests are welcome. The goal is to stay small and readable, so please keep new features to a few lines each.