Summary

Choosing a model becomes a query instead of a memory exercise.

search_models filters the registry by price ceiling, context length and capability, so the cheapest model meeting a hard requirement is one call rather than a browse through a pricing page. Completions then go wherever you picked, with provider routing and parameter overrides supplied per request. Four tools, and none of them a framework: this routes, it does not orchestrate.

What it is

Four tools over OpenRouter.ai: one sends messages to a named model, three answer questions about the registry — search it, describe one model, or check that an ID is valid.

What you get

Chat completions against a specified model, with temperature and max_tokens overridden per call
Provider routing passed alongside the request rather than fixed at startup
Registry search filtered by capabilities, provider, minimum and maximum context length, and prompt or completion price ceilings
Complete metadata for a single model ID: pricing, context limits and capabilities
A boolean check that a model ID exists, before another call depends on it

Requirements

An OpenRouter account and OPENROUTER_API_KEY. Four further environment variables hold defaults the calls fall back to: OPENROUTER_DEFAULT_MODEL, OPENROUTER_MAX_TOKENS, OPENROUTER_PROVIDER_QUANTIZATIONS and OPENROUTER_PROVIDER_IGNORE.

Setup effort

One command plus a key — pnpm install @mcpservers/openrouterai, then supply credentials