An open protocol from Anthropic (Nov 2024) for connecting AI agents to external tools, data sources, and prompts.
MCP cluster
MCP for your API.
Model Context Protocol (MCP) is the emerging standard for letting AI agents — Claude, Codex, Cursor, Antigravity, Cline — call your API as native tools instead of constructing HTTP requests by hand. If your API has an OpenAPI spec, you already have most of what an MCP server needs.
Completely free for 30 days. No credit card required.
01Details
What MCP actually is
Each MCP server exposes a list of tools (functions the agent can call), resources (data the agent can read), and prompts (templates the agent can invoke).
Transports: stdio (local) and Streamable HTTP (remote). Clients pick whichever fits.
Spec: https://modelcontextprotocol.io — already adopted by Claude Desktop, ChatGPT, VS Code (Copilot), OpenAI Codex, Cursor, Google Antigravity, Cline, Windsurf, and dozens of community clients.
02Details
MCP vs OpenAPI — which do you ship?
OpenAPI describes the wire contract: paths, parameters, schemas, security. Built for HTTP clients (humans + SDKs).
MCP describes the tool contract: callable functions with typed inputs/outputs. Built for AI agents.
They are not alternatives — most teams ship both. OpenAPI for SDKs and docs; MCP for agents.
An MCP server typically wraps an existing API. If you have OpenAPI, generating the MCP server is a near-mechanical translation.
03Details
Why ship MCP now (the early-adopter case)
Coding agents are increasingly the first integrator of your API. Whoever ships an MCP server first owns the canonical agent integration.
MCP cuts the integration surface from 'write an HTTP client' to 'declare a tool' — agents wire up correctly on the first call instead of after 5 hallucinations.
The official MCP registry (https://registry.modelcontextprotocol.io) is a discoverable channel. Listing early is cheap distribution.
If your customers use Claude / Cursor / Codex, your MCP server gets one-click install in their tool config.
04Details
What an MCP server for an API looks like (minimum viable)
A small Node or Python program that uses the MCP SDK (modelcontextprotocol/sdk).
One tool per OpenAPI operation (e.g. create_user, list_messages), with the OpenAPI request body becoming the MCP tool input schema.
Auth: the MCP server forwards the user's API key or OAuth token to the underlying HTTP API.
Distribution: published to npm or PyPI as your-api-mcp. Users add a 4-line block to their MCP client config.
05Details
How Bloom fits
Bloom's pipeline emits TypeScript and Python SDKs from OpenAPI today; MCP-server emission is in preview alongside the same pipeline.
Preview output: an npm-shaped package (@your-co/mcp) with one tool per OpenAPI operation, derived from the same schemas the SDKs use.
Stdio transport first; Streamable HTTP variant for hosted MCP endpoints is on the roadmap behind a single --port flag.
Free for 30 days; no credit card. Reach out if MCP generation is a launch requirement and we'll confirm timing.