The ranking is crowdsourced — Elo ratings and win rates from designarena.ai — and availability is checked against OpenRouter before a prompt goes out, so a model that tops the board but is not carried gets skipped rather than failing the call. query_specific_model accepts either naming scheme, which means a Design Arena name read off the leaderboard can be passed straight through without translating it into an OpenRouter ID.
A routing layer covering 4 tools, reading the designarena.ai crowdsourced rankings and sending prompts through OpenRouter.
- The current ranking as data: get_best_design_model returns the top-rated design model with its OpenRouter availability, and get_leaderboard pages through the full rankings with Elo ratings, win rates and a category filter.
- Prompting that follows the ranking: query_design_model selects the best available model automatically and skips any that OpenRouter does not carry.
- An explicit override: query_specific_model accepts either an OpenRouter model ID or a Design Arena model name, with temperature and max_tokens.
An OpenRouter account and its key in OPENROUTER_API_KEY, which query_design_model and query_specific_model require.
One command plus a key — OPENROUTER_API_KEY=sk-or-... npx -y mcp-bench-router, then supply credentials