Openrouter

What the Openrouter MCP server does

The Openrouter MCP server gives MCP clients access to model metadata from the OpenRouter catalog. Its primary use is model discovery and selection rather than sending prompts to a model. Results include normalized pricing per 1 million tokens, context-window size, modality, and, where available, additional model metadata.

The catalog contains roughly 340 models from major providers. Agents can search the catalog, restrict results to free models, require a minimum context window, or filter by modality such as image support. Results from model listing are ordered by completion price, from lowest to highest.

How it works

The project supports two connection patterns. A hosted MCP endpoint is available at https://gateway.pipeworx.io/openrouter/mcp. A local package can be started with npx -y @pipeworx/mcp-openrouter, communicating with the client through standard input and output.

The hosted pack endpoint lists the three OpenRouter-specific tools and also includes shared Pipeworx gateway tools. These shared tools can route questions across the wider Pipeworx catalog. The standalone process exposes only the OpenRouter pack tools. The full Pipeworx gateway is an alternative hosted endpoint that lists all available packs directly.

Without an MCP client, the hosted tools can also be called over HTTP. The documented endpoint for listing models is POST https://gateway.pipeworx.io/v1/tools/openrouter_list_models with a JSON request body.

Setup and configuration

For a hosted connection, add the pack endpoint to an MCP client configuration:

{
  "mcpServers": {
    "openrouter": {
      "url": "https://gateway.pipeworx.io/openrouter/mcp"
    }
  }
}

For local stdio execution, use:

{
  "mcpServers": {
    "openrouter": {
      "command": "npx",
      "args": ["-y", "@pipeworx/mcp-openrouter"]
    }
  }
}

No environment variables or OpenRouter API key are specified for either setup.

Tools and capabilities

• list_models searches the catalog and supports search text, free-only filtering, minimum context, and modality filtering. It returns model IDs, pricing, context windows, and modality.
• get_model retrieves one model by ID, such as anthropic/claude-sonnet-4.5. Its details can include prompt and completion pricing, per-request, image, and web-search costs, supported parameters, tokenizer, and knowledge cutoff.
• compare_models compares two to five model IDs by pricing, context window, and modality. It is suited to questions such as whether one named model costs less than another.

The Openrouter MCP server uses normalized dollar-per-million-token pricing for comparable results. The hosted gateway also provides ask_pipeworx, which can select and invoke a relevant tool from the wider Pipeworx catalog, while the standalone package does not include that routing layer.

Limitations and notes

This package is focused on OpenRouter model information. The documented tools do not describe prompt execution, completion generation, or model deployment. Hosted pack connections expose additional shared schemas beyond the three pack tools, so the tool count may be higher than expected; the connection's initialize response is the authoritative scope for that session.

The catalog size is approximate and may change. Model IDs must be supplied when requesting details or comparisons, and comparisons accept between two and five models. The standalone mode avoids the gateway round trip but does not provide the gateway's shared meta-tools or cross-catalog routing.