mcp-gen

Point it at an OpenAPI spec and it writes the MCP server. A 300-endpoint API becomes three tools and about 800 tokens of context, instead of 300 tool definitions the model has to read before it can do anything.

What mcp-gen Does

mcp-gen takes any OpenAPI specification and writes a TypeScript MCP server you can deploy. Feed it a spec URL or file, it outputs a deployable server with type-safe tools, credential management, and audit logging.

OpenAPI Spec → mcp-gen sync → Deployable MCP Server

The Token Problem

Most MCP servers register every tool upfront. When an LLM connects via tools/list, it receives the full catalog: every tool name, description, and JSON Schema input definition, injected into its context window. A hand-built Stripe MCP with 300 tools can burn 50-80K tokens before the user even asks a question.

This is why developers are ditching MCPs for raw CLI wrappers. But CLIs lose credential vaulting, access control, audit trails, and rate limiting: everything enterprises need.

Three Generation Modes

1. Generic mode: three tools, any API size

Registers exactly three tools regardless of whether the API has 10 or 1,000 endpoints.

Tool	Purpose	Input Schema
{name}_list_schemas	Discover endpoints by category or search	{ category?, search? }
{name}_get_schema	Get detailed params/body schema for one endpoint	{ operation }
{name}_api_call	Call any endpoint with auto-routed params	{ operation, params?, body? }

The LLM discovers progressively: browse → inspect → call. All endpoint metadata lives server-side in endpoint_map.json, never sent to the LLM until requested.

~800 tokens in the tools/list response. Always. Whether it's Petstore (20 endpoints) or Stripe (300+).

2. Per-endpoint mode: one tool per endpoint

One tool per API endpoint. Each tool has a typed input schema derived from the OpenAPI spec.

Best for small, focused APIs where every endpoint matters and you want the LLM to see the full surface area upfront.

~200-400 tokens per tool. A 20-endpoint API runs ~5K tokens. A 100-endpoint API runs ~30K+.

3. Curated mode: hand-picked, with a fallback

You choose 5-15 high-value endpoints as direct tools. Everything else is accessible through a fallback api_call tool.

curated:
  tools:
    - operation_id: createCustomer
    - operation_id: getInvoice
    - operation_id: createPaymentIntent
  fallback:
    enabled: true
  explorer:
    enabled: true

~2-4K tokens for 10 curated tools + fallback + explorer.

Token Usage Comparison

API Size	Hand-Built	Per-Endpoint	Curated (10)	Generic
20 endpoints	~8-12K	~5-8K	~3-4K	~800
50 endpoints	~20-30K	~15-20K	~3-4K	~800
100 endpoints	~40-60K	~30-40K	~3-4K	~800
300 endpoints	~100-150K	~80-120K	~3-4K	~800

Estimated based on typical tool description + JSON Schema sizes.

Unified Parameter Schema

Most MCP servers force the LLM to understand HTTP routing: path params, query strings, headers. mcp-gen flattens everything into a single params object.

What other MCPs do:

path: { petId: "123" }
query: { status: "available" }
headers: { "X-API-Key": "sk-..." }

What mcp-gen does:

params: { petId: "123", status: "available" }
// Routing handled server-side. Credentials from vault.

What You Get Beyond Token Savings

Credential vaulting · Secrets pulled from Vault at runtime, never exposed to the LLM

Audit logging · Every tool invocation logged with the identity behind it

Multi-tenant support · Customer ID from JWT, isolated credential scopes

Type-safe inputs · Zod validation on every parameter

Deployable to Statio · One command to Cloud Run with gateway routing and billing

Connect your first server.

Three MCP servers and 5,000 calls a month, free and without a card. If you outgrow it the next plan is $19, and it stays $19 as the team grows.

Start free

Keys are encrypted at rest and never handed to an agent
Cancel anytime. You drop to Free, nothing is deleted
No card required, and no way to spend money on Free

How Statio handles credentials →