openrouter-task2model

MCP Server to find the best AI model for your task. Searches 300+ OpenRouter models.

Installation

Claude Code

claude mcp add openrouter-task2model -e OPENROUTER_API_KEY=sk-or-... -- npx -y openrouter-task2model

Manual Configuration

{
  "mcpServers": {
    "openrouter-task2model": {
      "command": "npx",
      "args": ["-y", "openrouter-task2model"],
      "env": {
        "OPENROUTER_API_KEY": "sk-or-..."
      }
    }
  }
}

Design Philosophy

The AI decides which model to use, not the embeddings.

Price limits, age filters, and semantic search narrow down 300+ models to ~100 candidates. From there, the AI uses its own knowledge to decide which models to benchmark. The ranking within those 100 is not a quality indicator - it just helps reduce the search space.

Think of it as: price + date + embeddings filter, AI selects.

Tools

task2model

Find models for a task. Returns top 100 model IDs.

{ "task": "Build a coding assistant with tool use" }

Output:

{
  "task": "Build a coding assistant with tool use",
  "models": ["anthropic/claude-sonnet-4", "openai/gpt-4o", "..."],
  "count": 100,
  "price_range": "$0.10-$50/1M",
  "note": "Ranked by description similarity. Does not predict actual task performance. Benchmark before production use."
}

Options

| Parameter | Default | Description | |-----------|---------|-------------| | hard_constraints.max_age_days | 365 | Filter models older than N days | | hard_constraints.max_price_per_1m | - | Max price per 1M tokens | | hard_constraints.required_parameters | - | e.g. ["tools", "structured_outputs"] | | hard_constraints.input_modalities | - | e.g. ["text", "image"] | | hard_constraints.providers | - | e.g. ["anthropic", "openai"] | | result.limit | 100 | Number of models to return | | result.detail | names_only | names_only, minimal, standard, full |

get_model_profile

Get detailed info for a specific model.

{ "model_id": "google/gemini-2.5-flash" }

sync_catalog

Refresh the model catalog cache.

{ "force": true }

Known Limitations

Semantic ranking does NOT predict performance.

The ranking is based on model descriptions (marketing text), not benchmarks. Example:

| Model | Semantic Rank | Actual Performance | |-------|---------------|-------------------| | GPT-5.1-Codex-Mini | #6 | Poor | | Gemini 2.5 Flash | #77 | Best |

However, both models are in the top 100 candidates. The AI can pick either one to test - ranking within the 100 is not a quality signal. The AI should use its own judgment to select from the pool.

Environment Variables

| Variable | Description | |----------|-------------| | OPENROUTER_API_KEY | Required for semantic search | | CACHE_TTL_MS | Cache TTL (default: 600000) |

License

MIT