mcpField
O
openapi-dynamic

openapi-dynamic

Local
github.com/mayorandrewDeveloper Tools1Updated 29w ago

About openapi-dynamic

openapi-dynamic connects MCP clients and command-line users to APIs described by OpenAPI specs, generating tools to explore and call multiple APIs. It stands out with varied authentication flows, request previews, and response filtering.

Key features

  1. 01Load local or remote OpenAPI specs
  2. 02Browse APIs, endpoints, and schemas
  3. 03Call endpoints and preview requests
  4. 04Support API key, Basic, Bearer, and OAuth2 auth
  5. 05Filter responses with JSONPath
  6. 06Upload files and retry rate-limited requests

Use cases

  1. 01Connect multiple APIs through one MCP server
  2. 02Inspect endpoint parameters before making a request
  3. 03Preview URLs, headers, and bodies before sending
  4. 04Filter large responses or upload files

MCP support

As an MCP server, it provides tools to list APIs, inspect endpoints and schemas, and preview or execute endpoint requests.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
"mcpServers": {
"mayorandrew-openapi-dynamic": {
"command": "npx",
"args": [
"-y",
"openapi-dynamic-mcp"
]
}
}
}
Run commandnpx -y openapi-dynamic-mcp

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
npmopenapi-dynamic-mcp0.3.3

Tools

No tool list yet

Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.

Details

From the project README

openapi-dynamic-mcp

Connect AI clients to OpenAPI APIs quickly, with one MCP server, direct CLI access, and built-in auth support.

NPM Version License Node.js Version

Table of Contents

Overview

openapi-dynamic-mcp lets MCP clients and shell users work with OpenAPI APIs without writing custom glue code for each service. Point it at one or more OpenAPI specs, then list APIs, inspect endpoints, authenticate, and make requests through a consistent interface.

It is designed for common user workflows:

  • Connect multiple APIs through one MCP server
  • Use local specs or hosted specUrl definitions
  • Work with OpenAPI 3.0, 3.1, and Swagger 2.0
  • Handle API key, bearer, basic, and OAuth2 auth
  • Reuse stored tokens across sessions
  • Filter large responses down to the fields you need
  • Preview requests safely before sending them

Highlights

  • Get from spec to usable tools fast: start from a YAML config and immediately browse endpoints or call them from MCP or the CLI.
  • Authenticate the way your API expects: supports API keys, bearer/basic auth, and OAuth2 client credentials, password, device code, and auth code with PKCE.
  • Avoid repeated sign-in work: store tokens once and reuse them later.
  • Handle interactive OAuth cleanly: device-code and browser-based auth return instructions an agent can present to the user.
  • Keep responses focused: project large outputs with JSONPath selectors.
  • Inspect before you send: use dry runs to preview request shape without network I/O.
  • Upload files when needed: supports multipart form uploads and raw binary bodies.
  • Stay resilient against rate limits: configurable retries for 429 Too Many Requests.

Requirements

  • Node.js 20+

Quick Start

Run the MCP server directly:

npx -y openapi-dynamic-mcp@latest --config ./config.yaml

Minimal config:

version: 1
apis:
  - name: pet-api
    specPath: ./pet-api.yaml

You can also point at a remote spec:

version: 1
apis:
  - name: pet-api
    specUrl: https://api.example.com/openapi.json

Configuration

Add each API you want to use under apis. Each entry can point to a local spec file or a remote spec URL.

version: 1

apis:
  - name: pet-api
    specPath: ./pet-api.yaml
    # specUrl: https://api.example.com/openapi.yaml
    baseUrl: https://api.example.com/v1
    timeoutMs: 30000
    headers:
      X-Client: openapi-dynamic-mcp
    retry429:
      maxRetries: 2
      baseDelayMs: 250
      maxDelayMs: 5000
      jitterRatio: 0.2
      respectRetryAfter: true
    oauth2Schemes:
      OAuthCC:
        tokenUrl: https://auth.example.com/oauth2/token
        scopes: [read:pets, write:pets]
        tokenEndpointAuthMethod: client_secret_basic
      UserAuth:
        authMethod: device_code
        deviceAuthorizationEndpoint: https://auth.example.com/oauth/device
        pkce: true

Common options:

  • name: the API name shown in MCP and CLI commands
  • specPath or specUrl: where to load the OpenAPI spec from
  • baseUrl: override the server URL from the spec
  • headers: headers to send on every request
  • timeoutMs: default request timeout
  • retry429: retry behavior for rate-limited APIs
  • oauth2Schemes: per-scheme OAuth settings when the spec defines OAuth security
Per-Scheme OAuth2 Configuration

Use oauth2Schemes when an API defines one or more OAuth2 security schemes and you want to set token URLs, scopes, or interactive auth preferences for a specific scheme.

The scheme name must match the name in the OpenAPI spec. Common options are:

  • tokenUrl
  • scopes
  • tokenEndpointAuthMethod
  • authMethod
  • deviceAuthorizationEndpoint
  • pkce

If you only need credentials, environment variables are often enough. Use oauth2Schemes when you want reusable config checked into the project.

Client Setup

Claude Desktop / Claude Code
{
  "mcpServers": {
    "openapi": {
      "command": "npx",
      "args": [
        "-y",
        "openapi-dynamic-mcp@latest",
        "--config",
        "/absolute/path/to/config.yaml"
      ],
      "env": {
        "PET_API_BASE_URL": "http://localhost:3000"
      }
    }
  }
}
Cursor
{
  "mcpServers": {
    "openapi": {
      "command": "npx",
      "args": [
        "-y",
        "openapi-dynamic-mcp@latest",
        "--config",
        "/absolute/path/to/config.yaml"
      ]
    }
  }
}

CLI

Server mode is available as either the root command or the explicit serve subcommand:

openapi-dynamic-mcp --config ./config.yaml
openapi-dynamic-mcp serve --config ./config.yaml

Use the CLI when you want the same API access outside your MCP client, for scripting, debugging, or auth setup.

Tool Commands

Every MCP tool is also available as a CLI subcommand that accepts one JSON object and emits JSON output:

openapi-dynamic-mcp list_apis --config ./config.yaml --input '{}'
openapi-dynamic-mcp list_api_endpoints --config ./config.yaml --input '{"apiName":"pet-api"}'
openapi-dynamic-mcp get_api_endpoint --config ./config.yaml --input '{"apiName":"pet-api","endpointId":"listPets"}'
openapi-dynamic-mcp get_api_schema --config ./config.yaml --input '{"apiName":"pet-api","pointer":"/info"}'
openapi-dynamic-mcp make_endpoint_request --config ./config.yaml --input '{"apiName":"pet-api","endpointId":"listPets","dryRun":true}'

Shared flags:

  • --input <json>: JSON object with command arguments
  • --fields <jsonpath>: repeatable selector for filtering successful output
  • --describe: print the command schema and help metadata
  • --auth-file <path>: override the auth-store path
Auth Command

Use auth to pre-authenticate one configured security scheme and persist its token for later MCP or CLI calls:

openapi-dynamic-mcp auth --config ./config.yaml --api pet-api --scheme OAuthCC
openapi-dynamic-mcp auth --config ./config.yaml --api pet-api --scheme ApiKeyAuth --token secret

For API key and bearer auth, --token provides the secret directly. For OAuth2, the command uses your configured credentials, completes the flow, and stores the result for later use.

Authentication

Supported authentication types:

  • API key
  • HTTP bearer
  • HTTP basic
  • OAuth2 client credentials
  • OAuth2 password grant (ROPC)
  • OAuth2 device code
  • OAuth2 authorization code with PKCE

Typical auth flow:

  1. Provide credentials through environment variables or config
  2. Run auth once if the scheme needs a stored token
  3. Reuse that token from MCP or the CLI until it expires or changes
Auth Store

By default, tokens are stored beside your config file in:

.openapi-dynamic-mcp-auth.json

You can override that path with either:

  • --auth-file
  • OPENAPI_DYNAMIC_MCP_AUTH_FILE

This makes repeated API use much smoother, especially for MCP clients that need to reconnect often.

Interactive OAuth2 Flows

When a request needs user interaction, the tool returns structured guidance that an MCP agent can relay to the user. Typical device-code output looks like:

{
  "status": "authorization_required",
  "method": "device_code",
  "message": "User authorization required. Ask the user to visit the URL and enter the code.",
  "verificationUri": "https://auth.example.com/device",
  "userCode": "ABCD-1234",
  "instruction": "After the user confirms, call this endpoint again."
}

This is especially useful for MCP agents because the auth step becomes a normal part of the user workflow instead of a dead-end error.

Environment Variables

Environment variables are useful for secrets, base URL overrides, and CI setups.

Names are derived from normalized API and scheme names:

  • Uppercase
  • Non-alphanumeric characters become _
  • Repeated _ are collapsed
  • Leading and trailing _ are removed

Examples:

  • pet-api -> PET_API
  • OAuth2 -> OAUTH2
API-Level Variables
  • <API>_BASE_URL
  • <API>_HEADERS as a JSON object string
  • OPENAPI_DYNAMIC_MCP_AUTH_FILE
API Key
  • <API>_<SCHEME>_API_KEY
HTTP Auth
  • <API>_<SCHEME>_TOKEN
  • <API>_<SCHEME>_USERNAME
  • <API>_<SCHEME>_PASSWORD
OAuth2
  • <API>_<SCHEME>_ACCESS_TOKEN
  • <API>_<SCHEME>_CLIENT_ID
  • <API>_<SCHEME>_CLIENT_SECRET
  • <API>_<SCHEME>_TOKEN_URL
  • <API>_<SCHEME>_SCOPES as a space-delimited list
  • <API>_<SCHEME>_TOKEN_AUTH_METHOD as client_secret_basic or client_secret_post
  • <API>_<SCHEME>_USERNAME
  • <API>_<SCHEME>_PASSWORD
  • <API>_<SCHEME>_AUTH_METHOD as device_code or authorization_code
  • <API>_<SCHEME>_DEVICE_AUTHORIZATION_ENDPOINT
  • <API>_<SCHEME>_REDIRECT_PORT
  • <API>_<SCHEME>_PKCE as true or false

Useful behaviors:

  • _ACCESS_TOKEN bypasses OAuth grant flows entirely.
  • _AUTH_METHOD forces device_code or authorization_code when both are possible.
  • _USERNAME and _PASSWORD are used for both HTTP basic auth and OAuth password grant, depending on the security scheme.

Working with Responses

JSONPath Filtering

CLI --fields and MCP fields: string[] let you keep only the parts of a successful response you care about. This is helpful when specs or payloads are too large to inspect comfortably.

Examples:

openapi-dynamic-mcp list_apis --config ./config.yaml --fields '$.apis[*].name'
openapi-dynamic-mcp get_api_endpoint --config ./config.yaml --input '{"apiName":"pet-api","endpointId":"listPets"}' --fields '$.responses'

Selectors support quoted member escaping, array indexes, and wildcards.

Large Schema Warnings

get_api_schema adds a _sizeWarning advisory field when the response is very large, prompting you to narrow the JSON Pointer.

Dry Runs

make_endpoint_request supports dryRun: true so you can confirm the URL, headers, auth, and serialized body before sending a real request.

Files and Binary Data

make_endpoint_request supports both multipart/form-data and raw binary uploads.

Each file entry must provide exactly one content source: base64, text, or filePath.

{
  "name": "avatar.png",
  "contentType": "image/png",
  "filePath": "/absolute/path/to/avatar.png"
}

Multipart example:

{
  "apiName": "pet-api",
  "endpointId": "uploadProfile",
  "contentType": "multipart/form-data",
  "body": {
    "description": "A photo of Fido"
  },
  "files": {
    "profileImage": {
      "name": "fido.jpg",
      "contentType": "image/jpeg",
      "filePath": "/Users/local/images/fido.jpg"
    }
  }
}

Raw binary example:

{
  "apiName": "pet-api",
  "endpointId": "uploadRaw",
  "contentType": "application/octet-stream",
  "files": {
    "body": {
      "filePath": "/Users/local/data.bin"
    }
  }
}

MCP Tools

These five tools cover the main user workflows:

ToolPurpose
list_apisList configured APIs
list_api_endpointsSearch or paginate endpoints in one API
get_api_endpointInspect endpoint metadata, parameters, responses, and security
get_api_schemaReturn a schema object or JSON Pointer target from the spec
make_endpoint_requestPreview or execute an endpoint request

Development

npm install
npm run build
npm test

Useful commands:

npm run lint
npm run format
npm run test:watch

License

MIT

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.