Runapi AI MCP

What runapi-ai/mcp MCP server does

The runapi-ai/mcp MCP server exposes RunAPI through the Model Context Protocol. It gives an MCP client access to model discovery, endpoint inspection, runtime pricing, prompt examples, authenticated task operations, and account balance information. RunAPI presents a unified API across many models and providers, including image, video, music, audio, text-to-speech, and LLM-related operations described by the catalog.

The package is a client for RunAPI rather than a local inference or media-generation backend. Actual authenticated operations are sent to RunAPI, while discovery features can read from the package’s embedded catalog. This makes it suitable for assistants that need to compare available models and inputs before creating a task.

How it works

An assistant can begin with list_models, list_actions, or search_prompts. get_model_info returns the endpoint, input requirements, constraints, and current runtime pricing for a model slug. If a model is available through multiple endpoints, the service and action can be supplied to disambiguate it. check_pricing provides current pricing for a model and action rather than relying on memorized values.

For an authenticated workflow, create_task submits a RunAPI operation. It requires a caller-generated idempotency key and can either wait for completion or return immediately. Asynchronous work returned without waiting can be monitored with get_task. Reusing the same idempotency key is intended for safely retrying the same logical request; changing the request while reusing the key causes a conflict.

Setup and configuration

Install the local runapi-ai/mcp MCP server from npm with:

npx -y @runapi.ai/mcp

For manual MCP configuration, use npx as the command and -y, @runapi.ai/mcp as its arguments. The README also documents Claude Code commands that register the server at user or project scope, plus an init command for generating configurations for several clients.

Authenticated users can call the login tool to open a browser-based PKCE login flow. The resulting credentials are saved in ~/.config/runapi/config.json, which is shared with the RunAPI CLI. Headless or CI environments can provide RUNAPI_API_KEY before starting the MCP host.

A hosted endpoint is also available at https://mcp.runapi.ai/mcp. Remote clients do not need Node.js and can use OAuth where supported, or provide a RunAPI API key as a bearer credential. The hosted service exposes the documented business tools; local login is available through the local package.

Tools and capabilities

The runapi-ai/mcp MCP server provides these capabilities:

• Browse models with optional modality, service, and action filters.
• Inspect model inputs, endpoint support, constraints, and runtime prices.
• Group available actions by output modality.
• Search reusable prompts by text, modality, category, tags, model, or featured status.
• Create tasks synchronously or asynchronously with idempotency protection.
• Retrieve task status and the latest result payload.
• Authenticate through browser login and inspect account balance and spending metrics.

Limitations and notes

Catalog, pricing, and prompt-search access does not require an API key, but creating tasks, reading task status, and checking account balance do. Pricing is runtime data and should be queried when accuracy matters. A completed task may not be available before the client’s wait deadline; in that case, use get_task rather than creating a replacement. If a connection closes after submission, the original task may still be processing, so retry with the same idempotency key only when repeating the same request is intended.