Runapi AI MCP

What runapi-ai/mcp MCP server does

runapi-ai/mcp MCP server exposes RunAPI’s model catalog and API operations to MCP-compatible AI clients. The catalog covers image generation, video generation, music creation, text-to-speech, and other audio workflows, along with LLM endpoints. Agents can inspect available models, determine which actions they support, review input requirements, and retrieve runtime pricing before submitting work.

The package is a client for RunAPI rather than a local generation engine. Generated media and other operations are performed by RunAPI, while the MCP layer gives an assistant a consistent interface for discovery, authentication, task submission, and status retrieval.

How it works

Catalog, action, pricing, and prompt-search tools use an embedded catalog and can run without an API key. This makes it possible to discover model slugs and required parameters before signing in. Runtime pricing is returned for the selected model and action, helping an agent avoid relying on stale model information.

Authenticated operations use RunAPI credentials. The login tool opens a browser-based PKCE flow and stores the resulting key in ~/.config/runapi/config.json; the same shared configuration is used by runapi login. Headless environments can provide RUNAPI_API_KEY before starting the MCP host.

create_task accepts a caller-generated idempotency key. It can wait for an asynchronous operation to finish or return immediately when wait is false. When submitted without waiting, the returned task ID can be passed to get_task later. Reusing the same idempotency key is intended for retrying the same logical request, not for changing its inputs or creating a separate task.

Setup and configuration

For a local stdio installation, run the package through npx:

{
  "mcpServers": {
    "runapi": {
      "command": "npx",
      "args": ["-y", "@runapi.ai/mcp"]
    }
  }
}

Claude Code users can register it with claude mcp add runapi -s user -- npx -y @runapi.ai/mcp. Project scope is available when the configuration should be shared through a repository’s .mcp.json file. The package also provides init commands for Claude, Cursor, VS Code, Windsurf, and Roo configurations.

Remote clients can connect to https://mcp.runapi.ai/mcp without installing Node.js. OAuth can be used when supported by the client, or a RunAPI API key can be configured as the bearer credential. Local login is available with the package; the hosted endpoint exposes the business tools.

Tools and capabilities

• list_models filters the embedded model catalog by modality, service, or action.
• get_model_info returns endpoint details, input constraints, and current runtime pricing for a model slug.
• list_actions groups available RunAPI actions by output modality.
• check_pricing returns runtime pricing for a model, service, and action combination.
• search_prompts finds reusable prompt examples using modality, category, tags, text, model, featured status, and pagination.
• login starts browser authentication and saves shared local credentials.
• create_task submits an authenticated RunAPI operation and can optionally poll for completion.
• get_task retrieves the latest status and payload for an existing task.
• check_balance reports account balance and spending metrics.

Limitations and notes

Discovery tools do not require authentication, but task creation, task retrieval, and balance checks do. The server depends on RunAPI for execution and does not provide a local model or generation backend. Asynchronous requests may still be processing after a wait deadline or connection interruption; use get_task with the task ID rather than submitting a replacement. Keep the original idempotency key and request inputs when retrying an uncertain submission.