Do I need an OpenAPI spec?

For most of these, yes. openapi-mcp-generator needs OpenAPI 3.0 or later; LiteLLM takes OpenAPI 3.0, 3.1 or Swagger 2.0; TrueFoundry takes OpenAPI 3.0 or 3.1; Azure API Management works from an API it manages. FastMCP can start from a FastAPI app, or you write the tools by hand. Kaiva Bridge also takes a cURL command, a Postman collection or a GraphQL endpoint.

How do I keep a large API from overwhelming the AI?

Choose which operations become tools. Each tool here can: x-mcp and --default-include in openapi-mcp-generator, curation and tag filters in Speakeasy, route maps in FastMCP, allowed_tools in LiteLLM, operation selection in Azure API Management and TrueFoundry (up to 200 tools per server), and per-tool switches in Kaiva Bridge. FastMCP's docs note that models do significantly better with curated servers.

Which of these are open source?

openapi-mcp-generator (MIT), FastMCP (Apache 2.0), LiteLLM (MIT, apart from its enterprise features), and Speakeasy's generators and Gram (AGPL-3.0). Azure API Management, TrueFoundry and Kaiva Bridge are commercial services; Kaiva Bridge's CLI is open source under MIT.

What happened to Stainless's hosted MCP servers?

Stainless announced on 18 May 2026 that it is joining Anthropic and winding down its hosted products. Its docs say Stainless-hosted MCP servers are deprecated and point to self-hosting its Docker image instead.

COMPARISON

Tools that turn an API into an MCP server

Checked 19 Oct 2026, from each tool's own docs, repository and pricing pages.

There are four ways to turn an API into an MCP server. Generate code and run it yourself (openapi-mcp-generator, or Speakeasy's open-source generator); write the server with a framework that can start from your spec (FastMCP); add the API to a gateway you already run (LiteLLM, or Azure API Management if your APIs are there); or use a managed platform that builds and serves the server (Speakeasy's AI control plane, TrueFoundry's MCP Gateway, Kaiva Bridge). The first two give you code to own and run; the last two give you an address without running a server. Apidog is often listed too: it gives AI coding tools your API's documentation and tests MCP servers, rather than turning endpoints into tools.

Four ways to do it

Generate code, then run it

A command reads your OpenAPI spec and writes an MCP server you deploy and maintain like any other service.

openapi-mcp-generator, Speakeasy's generator

Write it with a framework

You write the server in code, and the framework can start you off with a tool for each endpoint of your spec.

FastMCP

Add it to a gateway you run

A gateway you already operate reads the spec and serves the API as tools, next to everything else behind it.

LiteLLM, Azure API Management

Use a managed platform

You give the platform your spec or source, choose the tools, and it serves the server at an address. Nothing to deploy.

Speakeasy's AI control plane, TrueFoundry MCP Gateway, Kaiva Bridge

At a glance

Tools that turn an API into an MCP server, checked 19 Oct 2026
Tool	Approach	Turns into tools	Who runs it	Price
openapi-mcp-generator	ApproachGenerates code you run	Turns into toolsAn OpenAPI 3.0 or later spec, YAML or JSON	Who runs itYou	PriceFree, MIT
Speakeasy	ApproachGenerates code you run, or a managed platform	Turns into toolsOpenAPI 3.0 and 3.1 (3.2 in part); TypeScript functions; existing MCP servers	Who runs itYou (generated server) or Speakeasy	PriceGenerator open source (AGPL-3.0). Platform: Enterprise, tailored pricing
FastMCP	ApproachA framework you write in	Turns into toolsAn OpenAPI spec or a FastAPI app as a starting point; any Python you write	Who runs itYou, or Prefect's FastMCP Gateway	PriceFramework free (Apache 2.0). FastMCP Gateway: free for one developer, then $35 per developer a month plus usage
LiteLLM	ApproachA gateway you run	Turns into toolsOpenAPI 3.0, 3.1 or Swagger 2.0, by file path or URL	Who runs itYou	PriceFree, MIT (enterprise features by custom pricing)
Azure API Management	ApproachA gateway you run (in Azure)	Turns into toolsREST APIs managed in API Management, usually imported from OpenAPI	Who runs itAzure, in your subscription	PriceFrom $0.0658 an hour on the Developer tier (about $48 a month, East US list price)
TrueFoundry MCP Gateway	ApproachA managed platform	Turns into toolsOpenAPI 3.0 or 3.1, by URL or pasted in; up to 200 tools per server	Who runs itTrueFoundry	PriceFree for up to 3 users; Pro $25 per user a month
Apidog	ApproachAPI docs for AI coding tools; tests MCP servers	Turns into toolsDoes not turn endpoints into tools	Who runs itYou (npx)	PriceFree plan for up to 4 users
Kaiva Bridge	ApproachA managed platform	Turns into toolsREST APIs (OpenAPI, Swagger, a cURL command or a Postman collection), GraphQL, Postgres, MySQL, MariaDB and SQL Server, spreadsheets, documents, remote MCP servers	Who runs itKaiva	PriceFree: 10 servers and 200,000 tool calls a month. Pro: $19 a month

Each tool, and when it fits

openapi-mcp-generator

github.com/harsha-iiiv/openapi-mcp-generator

An open-source command-line tool that generates a Node.js and TypeScript MCP server from an OpenAPI spec, given as a file or a URL.

Transports: stdio by default; also SSE, Streamable HTTP and a Cloudflare Workers target.
Credentials: API key, Bearer, Basic or OAuth2 credentials for your API come from environment variables. --header-passthrough lets each user's key come from the client's headers. Sign-in for incoming clients is not documented.
Large APIs: Choose operations with the x-mcp extension in the spec and the --default-include option.
Release: 4.1.0, 24 Sep 2026, on npm.

Fits when you want the server's code in your own repository, you will run it yourself, locally or deployed, and your API has an OpenAPI 3 spec.

Speakeasy

www.speakeasy.com

Speakeasy describes itself as an AI control plane for companies rolling out agents, built around an MCP gateway. Its OpenAPI generators, which also generate a standalone TypeScript MCP server, were open-sourced under AGPL-3.0 on 17 Sep 2026, and Gram is the open-source stack behind its platform.

Sign-in: On the gateway: OAuth for organization members, managed OAuth with dynamic client registration (a Pro and Enterprise feature, per its docs), pre-obtained tokens and client credentials.
Large APIs: Choose which tools are exposed and filter by tag on the gateway. A generated server groups tools into scopes picked when it starts.
Open source: The generators and Gram are AGPL-3.0, with a commercial licence available.
Pricing: The pricing page lists one plan, Enterprise, priced Tailored.

Fits when you are rolling out agents across a company and want one gateway with company sign-in and roles, or you ship an official MCP server to your API's customers and regenerate it from your spec in CI.

FastMCP

gofastmcp.com

An open-source Python framework from Prefect for writing MCP servers. FastMCP.from_openapi() and FastMCP.from_fastapi() turn a spec or a FastAPI app into tools you can then reshape in code. Prefect's FastMCP Gateway hosts FastMCP servers.

Sign-in: Off by default. You add a provider: OAuth with dynamic client registration, an OAuth proxy for providers without it, or token verification. FastMCP Gateway includes managed OAuth.
Large APIs: Every endpoint becomes a tool by default; route maps include or leave out routes. FastMCP's docs say models do significantly better with curated servers, especially for complex APIs.
Release: 4.1.1, 18 Oct 2026, on PyPI.
Hosting: FastMCP Gateway's free Personal plan covers one developer and 200 servers; Developer is $35 per developer a month plus usage; Enterprise is custom.

Fits when you write Python, want to curate the tools or add your own logic around the API, and will run the server yourself or on FastMCP Gateway.

LiteLLM

docs.litellm.ai/docs/mcp_openapi

An open-source gateway for models and MCP servers that you run yourself. An OpenAPI spec added under mcp_servers in its config.yaml becomes a server, with one tool per operationId.

Credentials: Clients use LiteLLM keys. Calls to your API carry one configured credential: none, an API key, a bearer token, basic, an Authorization header or OAuth2.
Large APIs: allowed_tools and disallowed_tools choose what is offered.
Release: 1.106.0, 18 Oct 2026, on PyPI.
Licence: MIT, except the enterprise directory; enterprise is custom pricing with a 30-day trial.

Fits when you already run LiteLLM in front of your models and want your APIs as tools behind the same keys, logs and limits.

Azure API Management

learn.microsoft.com/azure/api-management/export-rest-mcp-server

Azure's API gateway can expose a REST API it already manages as a remote MCP server. You choose which operations become tools.

Limits: Tools only, no resources or prompts. Not available in workspaces. Only HTTP-compatible APIs.
Sign-in: A subscription key, or Microsoft Entra ID tokens. The credential manager can add OAuth 2.0 tokens on calls to your API.
Tiers: Developer, Basic, Standard and Premium and their v2 versions, and the self-hosted gateway.
Price: Developer $0.0658 an hour; Basic v2 $0.20548 an hour plus $0.03 per 10,000 calls (East US list prices).

Fits when your APIs already live in Azure API Management and your people sign in with Microsoft Entra ID.

TrueFoundry MCP Gateway

www.truefoundry.com/docs/ai-gateway/mcp/openapi-mcp-server

A managed AI gateway for models and MCP servers. Import from OpenAPI Spec builds an MCP server from your spec, and you select the tools to expose.

Protocol: Streamable HTTP at MCP version 2025-06-18; tools only.
Sign-in: TrueFoundry API keys, tokens from your identity provider, or TrueFoundry OAuth for IDEs. Calls to your API can use each user's own OAuth sign-in or API key, or pass their token through.
Large APIs: Up to 200 tools per OpenAPI server (raised from 30 in July 2026).
Pricing: Developer $0 for up to 3 users and 10K requests and tool calls per user a month; Pro $25 per user a month with 20K per user, then $20 per 100K; Enterprise custom.

Fits when you want one managed gateway for models and MCP servers across a team, with each user signing in to the upstream API as themselves.

Apidog

docs.apidog.com/apidog-mcp-server

An API design and testing platform. Its MCP Server gives AI coding tools such as Cursor your API's documentation, from an Apidog project, published Apidog docs or an OpenAPI file, so they can write code against it. Its built-in MCP client tests and debugs MCP servers.

What it is not: Its documentation does not describe exposing your API's endpoints as callable tools.
Testing: The MCP client connects over stdio or Streamable HTTP with an API key, bearer, JWT, basic, digest or OAuth 2.0.
Release: apidog-mcp-server 0.0.18, 9 Oct 2026, on npm.

Fits when you want your coding AI to write code against your API from its documentation, or you need to test an MCP server you built.

Kaiva Bridge

kaiv.ai/bridge

A managed MCP cloud platform. You connect a source, Bridge builds the tools from it and serves the MCP server.

Sign-in: OAuth 2.1 sign-in with a Kaiva account on every plan, so Claude and ChatGPT need no key, or a gateway key limited to the servers you pick.
Large APIs: Tools from an API or a database start switched off; you switch on the ones your AI needs.
Credentials: Stored encrypted by Bridge, added to each call on the server side and never sent to the AI client.
Pricing: Free: 10 servers and 200,000 tool calls a month, no card. Pro: $19 a month for unlimited servers and 1,000,000 tool calls, then $0.04 per 1,000 calls. Enterprise by contract.

Fits when what your AI needs already exists as an API, a database, a spreadsheet or documents, you want a working address with no code and no deploy step, and you want Claude, ChatGPT, Cursor and VS Code to connect without setting up sign-in yourself.

Questions

Do I need an OpenAPI spec?: For most of these, yes. openapi-mcp-generator needs OpenAPI 3.0 or later; LiteLLM takes OpenAPI 3.0, 3.1 or Swagger 2.0; TrueFoundry takes OpenAPI 3.0 or 3.1; Azure API Management works from an API it manages. FastMCP can start from a FastAPI app, or you write the tools by hand. Kaiva Bridge also takes a cURL command, a Postman collection or a GraphQL endpoint.
How do I keep a large API from overwhelming the AI?: Choose which operations become tools. Each tool here can: x-mcp and --default-include in openapi-mcp-generator, curation and tag filters in Speakeasy, route maps in FastMCP, allowed_tools in LiteLLM, operation selection in Azure API Management and TrueFoundry (up to 200 tools per server), and per-tool switches in Kaiva Bridge. FastMCP's docs note that models do significantly better with curated servers.
Which of these are open source?: openapi-mcp-generator (MIT), FastMCP (Apache 2.0), LiteLLM (MIT, apart from its enterprise features), and Speakeasy's generators and Gram (AGPL-3.0). Azure API Management, TrueFoundry and Kaiva Bridge are commercial services; Kaiva Bridge's CLI is open source under MIT.
What happened to Stainless's hosted MCP servers?: Stainless announced on 18 May 2026 that it is joining Anthropic and winding down its hosted products. Its docs say Stainless-hosted MCP servers are deprecated and point to self-hosting its Docker image instead.

Sources

Read on 19 Oct 2026. Tools, prices and features change, so check these pages before you decide. If something here is out of date, email [email protected] and we will correct it.

Try Kaiva Bridge

Connect an API, a database, a spreadsheet or documents and get an MCP server address. Free to start, no card.

Start free Read the docs

Tools that turn an API into an MCP server

Four ways to do it

Generate code, then run it

Write it with a framework

Add it to a gateway you run

Use a managed platform

At a glance

Each tool, and when it fits

openapi-mcp-generator

Speakeasy

FastMCP

LiteLLM

Azure API Management

TrueFoundry MCP Gateway

Apidog

Kaiva Bridge

Questions

Sources

Try Kaiva Bridge

Side-by-side comparisons