Connect AI clients to OpenAPI APIs quickly, with one MCP server, direct CLI access, and built-in auth support.
- Overview
- Highlights
- Requirements
- Quick Start
- Configuration
- Client Setup
- CLI
- Authentication
- Environment Variables
- Working with Responses
- Files and Binary Data
- MCP Tools
- Development
- License
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
specUrldefinitions - Work with OpenAPI
3.0,3.1, and Swagger2.0 - 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
- 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.
- Node.js
20+
Run the MCP server directly:
npx -y openapi-dynamic-mcp@latest --config ./config.yamlMinimal config:
version: 1
apis:
- name: pet-api
specPath: ./pet-api.yamlYou can also point at a remote spec:
version: 1
apis:
- name: pet-api
specUrl: https://api.example.com/openapi.jsonAdd 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: trueCommon options:
name: the API name shown in MCP and CLI commandsspecPathorspecUrl: where to load the OpenAPI spec frombaseUrl: override the server URL from the specheaders: headers to send on every requesttimeoutMs: default request timeoutretry429: retry behavior for rate-limited APIsoauth2Schemes: per-scheme OAuth settings when the spec defines OAuth security
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:
tokenUrlscopestokenEndpointAuthMethodauthMethoddeviceAuthorizationEndpointpkce
If you only need credentials, environment variables are often enough. Use oauth2Schemes when you want reusable config checked into the project.
{
"mcpServers": {
"openapi": {
"command": "npx",
"args": [
"-y",
"openapi-dynamic-mcp@latest",
"--config",
"/absolute/path/to/config.yaml"
],
"env": {
"PET_API_BASE_URL": "http://localhost:3000"
}
}
}
}{
"mcpServers": {
"openapi": {
"command": "npx",
"args": [
"-y",
"openapi-dynamic-mcp@latest",
"--config",
"/absolute/path/to/config.yaml"
]
}
}
}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.yamlUse the CLI when you want the same API access outside your MCP client, for scripting, debugging, or auth setup.
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
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 secretFor 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.
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:
- Provide credentials through environment variables or config
- Run
authonce if the scheme needs a stored token - Reuse that token from MCP or the CLI until it expires or changes
By default, tokens are stored beside your config file in:
.openapi-dynamic-mcp-auth.json
You can override that path with either:
--auth-fileOPENAPI_DYNAMIC_MCP_AUTH_FILE
This makes repeated API use much smoother, especially for MCP clients that need to reconnect often.
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 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_APIOAuth2->OAUTH2
<API>_BASE_URL<API>_HEADERSas a JSON object stringOPENAPI_DYNAMIC_MCP_AUTH_FILE
<API>_<SCHEME>_API_KEY
<API>_<SCHEME>_TOKEN<API>_<SCHEME>_USERNAME<API>_<SCHEME>_PASSWORD
<API>_<SCHEME>_ACCESS_TOKEN<API>_<SCHEME>_CLIENT_ID<API>_<SCHEME>_CLIENT_SECRET<API>_<SCHEME>_TOKEN_URL<API>_<SCHEME>_SCOPESas a space-delimited list<API>_<SCHEME>_TOKEN_AUTH_METHODasclient_secret_basicorclient_secret_post<API>_<SCHEME>_USERNAME<API>_<SCHEME>_PASSWORD<API>_<SCHEME>_AUTH_METHODasdevice_codeorauthorization_code<API>_<SCHEME>_DEVICE_AUTHORIZATION_ENDPOINT<API>_<SCHEME>_REDIRECT_PORT<API>_<SCHEME>_PKCEastrueorfalse
Useful behaviors:
_ACCESS_TOKENbypasses OAuth grant flows entirely._AUTH_METHODforcesdevice_codeorauthorization_codewhen both are possible._USERNAMEand_PASSWORDare used for both HTTP basic auth and OAuth password grant, depending on the security scheme.
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.
get_api_schema adds a _sizeWarning advisory field when the response is very large, prompting you to narrow the JSON Pointer.
make_endpoint_request supports dryRun: true so you can confirm the URL, headers, auth, and serialized body before sending a real request.
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"
}
}
}These five tools cover the main user workflows:
| Tool | Purpose |
|---|---|
list_apis |
List configured APIs |
list_api_endpoints |
Search or paginate endpoints in one API |
get_api_endpoint |
Inspect endpoint metadata, parameters, responses, and security |
get_api_schema |
Return a schema object or JSON Pointer target from the spec |
make_endpoint_request |
Preview or execute an endpoint request |
npm install
npm run build
npm testUseful commands:
npm run lint
npm run format
npm run test:watchMIT