JitAPI is an MCP server for dynamic API discovery and execution, enabling Claude to interact with the OpenAPI specifications of arbitrary APIs. Its core idea is: instead of dumping hundreds of endpoints into the context, it uses semantic search and a dependency graph to present only what is needed, and then Claude plans and executes the calls. According to the official description, Stripe has 300+ endpoints and GitHub has 800+, loading the full specification into Claude's context wastes tokens and leads to hallucinations; writing a custom MCP server for each API used also does not scale. JitAPI solves this problem: register any OpenAPI specification once, then state your needs in natural language, and it will find the correct endpoints, resolve the dependencies between them, and let Claude execute the calls.
Main capabilities:
- Multi-API orchestration: Register multiple APIs and ask cross-API questions. JitAPI searches across all registered APIs, and Claude chains the calls together.
- Registration: Parses the OpenAPI specification, builds a dependency graph (which endpoints need data from which other endpoints), and creates searchable embeddings for all endpoints.
- Search: When the user asks a question, JitAPI embeds the query and finds the most relevant endpoints via cosine similarity.
- Expansion: The dependency graph supplements any prerequisite endpoints (for example, "you need to call GET /users first to get the user_id before you can POST /orders").
- Execution: Claude receives the endpoint schemas and makes the API calls, passing data between steps.
How to use: Add it to the Claude Code configuration (.mcp.json). The official description says no API key is required—JitAPI works out of the box using local embeddings. You can create .mcp.json in the project directory (or use ~/.claude.json for global access), or add it to the Claude Desktop configuration.
Embedding providers: JitAPI works out of the box using local embeddings (fastembed), with no API key required. For better search quality on large APIs, you can add a cloud embedding provider. Set the API key in the env block of the MCP configuration. The provider is automatically detected based on the available environment variables, with priority order: Voyage AI > OpenAI > Cohere > local.
Authentication: Configure API authentication after registration. Supported authentication types: bearer, api_key (custom header), and api_key_query (query parameter). The recommended method uses environment variables, so that secrets are never written to disk. Then tell Claude to use the environment variable. When using env_var, JitAPI reads the secret from the environment at request time—only the environment variable name is persisted, never the credentials themselves. You can also pass credentials directly (they will be stored in ~/.jitapi/auth.json with 0600 permissions).
Security notes: When using env_var, credentials are resolved at runtime and never touch the filesystem. When passing credentials directly, secrets are stored in plaintext JSON in ~/.jitapi/auth.json (file permissions 0600, directory 0700). For production use, prefer the env_var method.
Suitable scenarios: Suitable for scenarios that require Claude to interact with large or complex OpenAPI specifications, especially APIs with a huge number of endpoints (such as Stripe and GitHub), as well as tasks that require orchestrating multi-step workflows across multiple APIs. Through semantic search and a dependency graph, it can reduce context usage and lower the risk of hallucinations.
Limitations: The material does not provide more detailed information on error handling, performance benchmarks, the range of supported OpenAPI versions, concurrency limits, or platform support beyond the above.
