JitAPI

Point Claude at any API. JitAPI figures out which endpoints to call and in what order — automatically.

JitAPI is an MCP server that lets Claude interact with any API from its OpenAPI spec. Instead of dumping hundreds of endpoints into context, JitAPI uses semantic search and a dependency graph to surface only what's needed — then Claude plans and executes the calls.

https://github.com/user-attachments/assets/53f72f89-a41a-4a9c-a688-ec876ea05fbd

The Problem

Stripe has 300+ endpoints. GitHub has 800+. Loading the full spec into Claude's context wastes tokens and causes hallucinations. Writing a custom MCP server for every API you use doesn't scale.

JitAPI solves this: register any OpenAPI spec once, then ask for what you need in plain English. It finds the right endpoints, resolves dependencies between them, and lets Claude execute the calls.

Quick Start

pip install jitapi

Add to your Claude Code config (.mcp.json):

{
  "mcpServers": {
    "jitapi": {
      "command": "uvx",
      "args": ["jitapi"]
    }
  }
}

That's it. No API keys required — JitAPI uses local embeddings out of the box.

Then in Claude:

You: Register the GitHub API from https://raw.githubusercontent.com/github/rest-api-description/main/descriptions/api.github.com/api.github.com.json

Claude: ✓ Registered GitHub v3 REST API — 1,107 endpoints indexed

You: List my repos

Claude: [searches for "list repositories for authenticated user" → finds GET /user/repos → executes]
Here are your repositories: ...

Multi-API Orchestration

The killer feature: register multiple APIs and ask questions that span them. JitAPI searches across all registered APIs and Claude chains the calls.

You: Register the TMDB API and OpenWeatherMap API
Claude: ✓ Registered both APIs

You: Find the top popular movie on TMDB, then get the weather where it was filmed

Claude: [searches TMDB → GET /movie/popular → GET /movie/{id} for production locations
         → searches OpenWeather → GET /data/2.5/weather with the city]

The #1 popular movie is "Inception", filmed in Los Angeles.
Current weather in LA: 72°F, partly cloudy.

How It Works

Register API                          Ask a question
     │                                      │
     ▼                                      ▼
Parse OpenAPI spec               Embed query → vector search
     │                                      │
     ▼                                      ▼
Build dependency graph           Find relevant endpoints
     │                                      │
     ▼                                      ▼
Embed all endpoints              Expand with dependencies
     │                                      │
     ▼                                      ▼
Store in vector DB               Return schemas → Claude executes

1. Register — Parse an OpenAPI spec, build a dependency graph (which endpoints need data from which other endpoints), and create searchable embeddings for all endpoints

2. Search — When you ask a question, JitAPI embeds your query and finds the most relevant endpoints via cosine similarity

3. Expand — The dependency graph adds any prerequisite endpoints (e.g., "you need to call GET /users first to get the user_id for POST /orders")

4. Execute — Claude gets the endpoint schemas and makes the API calls, passing data between steps

MCP Tools

Setup

Claude Code

Create .mcp.json in your project directory (or ~/.claude.json for global access):

{
  "mcpServers": {
    "jitapi": {
      "command": "uvx",
      "args": ["jitapi"]
    }
  }
}

Claude Desktop

Add to your Claude Desktop config:

{
  "mcpServers": {
    "jitapi": {
      "command": "uvx",
      "args": ["jitapi"]
    }
  }
}

Embedding Providers

JitAPI works out of the box with local embeddings (fastembed) — no API keys needed. For better search quality on large APIs, you can add a cloud embedding provider:

Set the API key in your MCP config's env block:

{
  "mcpServers": {
    "jitapi": {
      "command": "uvx",
      "args": ["jitapi"],
      "env": {
        "VOYAGE_API_KEY": "your-key-here"
      }
    }
  }
}

Provider is auto-detected from available environment variables. Priority: Voyage AI > OpenAI > Cohere > local.

Authentication

Configure API authentication after registering. The recommended approach uses environment variables so secrets are never written to disk:

{
  "mcpServers": {
    "jitapi": {
      "command": "uvx",
      "args": ["jitapi"],
      "env": {
        "GITHUB_TOKEN": "ghp_...",
        "OPENWEATHER_API_KEY": "your-key-here"
      }
    }
  }
}

Then tell Claude to use the env var:

You: Set bearer auth for GitHub using env var GITHUB_TOKEN
Claude: [calls set_api_auth with auth_type="bearer", env_var="GITHUB_TOKEN"]
✓ Auth configured for github (from env var $GITHUB_TOKEN)

With env_var, JitAPI reads the secret from the environment at request time — only the env var name is persisted, never the credential itself.

You can also pass credentials directly (they'll be stored in ~/.jitapi/auth.json with 0600 permissions):

You: Set API key auth for OpenWeather with param name "appid"
Claude: [calls set_api_auth with auth_type="api_key_query", credential="...", param_name="appid"]
✓ Auth configured for openweather

Supported auth types: bearer, api_key_header, api_key_query, basic.

Security note: When using env_var, credentials are resolved at runtime and never touch the filesystem. When passing credential directly, secrets are stored as plaintext JSON at ~/.jitapi/auth.json (file permissions 0600, directory 0700). For production use, prefer the env_var approach.

Environment Variables

Development

git clone https://github.com/nk3750/jitapi.git
cd jitapi
pip install -e ".[dev]"
pytest
ruff check src/

License

MIT