OpenRouter MCP Server

A Model Context Protocol (MCP) server providing seamless integration with OpenRouter.ai's diverse model ecosystem. Access various AI models through a unified, type-safe interface with built-in caching, rate limiting, and error handling.

Features

• Model Access

  • Direct access to all OpenRouter.ai models

  • Automatic model validation and capability checking

  • Default model configuration support

• Performance Optimization

  • Smart model information caching (1-hour expiry)

  • Automatic rate limit management

  • Exponential backoff for failed requests

• Unified Response Format

  • Consistent ToolResult structure for all responses

  • Clear error identification with isError flag

  • Structured error messages with context

Related MCP server: OpenRouter MCP Multimodal Server

Installation

pnpm install @mcpservers/openrouterai

Configuration

Prerequisites

1. Get your OpenRouter API key from OpenRouter Keys

2. Choose a default model (optional)

Environment Variables

• OPENROUTER_API_KEY: Required. Your OpenRouter API key.

• OPENROUTER_DEFAULT_MODEL: Optional. The default model to use if not specified in the request (e.g., openrouter/auto).

• OPENROUTER_MAX_TOKENS: Optional. Default maximum number of tokens to generate if max_tokens is not provided in the request.

• OPENROUTER_PROVIDER_QUANTIZATIONS: Optional. Comma-separated list of default quantization levels to filter by (e.g., fp16,int8) if provider.quantizations is not provided in the request. (Phase 1)

• OPENROUTER_PROVIDER_IGNORE: Optional. Comma-separated list of default provider names to ignore (e.g., mistralai,openai) if provider.ignore is not provided in the request. (Phase 1)

• OPENROUTER_PROVIDER_SORT: Optional. Default sort order for providers ("price", "throughput", or "latency"). Overridden by provider.sort argument. (Phase 2)

• OPENROUTER_PROVIDER_ORDER: Optional. Default prioritized list of provider IDs (JSON array string, e.g., '["openai/gpt-4o", "anthropic/claude-3-opus"]'). Overridden by provider.order argument. (Phase 2)

• OPENROUTER_PROVIDER_REQUIRE_PARAMETERS: Optional. Default boolean (true or false) to only use providers supporting all specified request parameters. Overridden by provider.require_parameters argument. (Phase 2)

• OPENROUTER_PROVIDER_DATA_COLLECTION: Optional. Default data collection policy ("allow" or "deny"). Overridden by provider.data_collection argument. (Phase 2)

• OPENROUTER_PROVIDER_ALLOW_FALLBACKS: Optional. Default boolean (true or false) to control fallback behavior if preferred providers fail. Overridden by provider.allow_fallbacks argument. (Phase 2)

# Example .env file content
OPENROUTER_API_KEY=your-api-key-here
OPENROUTER_DEFAULT_MODEL=openrouter/auto
OPENROUTER_MAX_TOKENS=1024
OPENROUTER_PROVIDER_QUANTIZATIONS=fp16,int8
OPENROUTER_PROVIDER_IGNORE=openai,anthropic
OPENROUTER_PROVIDER_SORT=price
OPENROUTER_PROVIDER_ORDER='["openai/gpt-4o", "anthropic/claude-3-opus"]'
OPENROUTER_PROVIDER_REQUIRE_PARAMETERS=true
OPENROUTER_PROVIDER_DATA_COLLECTION=deny
OPENROUTER_PROVIDER_ALLOW_FALLBACKS=false

OPENROUTER_PROVIDER_QUANTIZATIONS=fp16,int8 OPENROUTER_PROVIDER_IGNORE=openai,anthropic

### Setup

Add to your MCP settings configuration file (`cline_mcp_settings.json` or `claude_desktop_config.json`):

```json
{
  "mcpServers": {
    "openrouterai": {
      "command": "npx",
      "args": ["@mcpservers/openrouterai"],
      "env": {
        "OPENROUTER_API_KEY": "your-api-key-here",
        "OPENROUTER_DEFAULT_MODEL": "optional-default-model",
        "OPENROUTER_MAX_TOKENS": "1024",
        "OPENROUTER_PROVIDER_QUANTIZATIONS": "fp16,int8",
        "OPENROUTER_PROVIDER_IGNORE": "openai,anthropic"
      }
    }
  }
}

## Response Format

All tools return responses in a standardized structure:

```typescript
interface ToolResult {
  isError: boolean;
  content: Array<{
    type: "text";
    text: string; // JSON string or error message
  }>;
}

Success Example:

{
  "isError": false,
  "content": [{
    "type": "text",
    "text": "{\"id\": \"gen-123\", ...}"
  }]
}

Error Example:

{
  "isError": true,
  "content": [{
    "type": "text",
    "text": "Error: Model validation failed - 'invalid-model' not found"
  }]
}

Available Tools

chat_completion

Sends a request to the OpenRouter Chat Completions API.

Input Schema:

• model (string, optional): The model to use (e.g., openai/gpt-4o, google/gemini-pro). Overrides OPENROUTER_DEFAULT_MODEL. Defaults to openrouter/auto if neither is set.

  • Model Suffixes: You can append :nitro to a model ID (e.g., openai/gpt-4o:nitro) to potentially route to faster, experimental versions if available. Append :floor (e.g., mistralai/mistral-7b-instruct:floor) to use the cheapest available variant of a model, often useful for testing or low-cost tasks. Note: Availability of :nitro and :floor variants depends on OpenRouter.

• messages (array, required): An array of message objects conforming to the OpenAI chat completion format.

• temperature (number, optional): Sampling temperature. Defaults to 1.

• max_tokens (number, optional): Maximum number of tokens to generate in the completion. Overrides OPENROUTER_MAX_TOKENS.

• provider (object, optional): Provider routing configuration. Overrides corresponding OPENROUTER_PROVIDER_* environment variables.

  • quantizations (array of strings, optional): List of quantization levels to filter by (e.g., ["fp16", "int8"]). Only models matching one of these levels will be considered. Overrides OPENROUTER_PROVIDER_QUANTIZATIONS. (Phase 1)

  • ignore (array of strings, optional): List of provider names to exclude (e.g., ["openai", "anthropic"]). Models from these providers will not be used. Overrides OPENROUTER_PROVIDER_IGNORE. (Phase 1)

  • sort ("price" | "throughput" | "latency", optional): Sort providers by the specified criteria. Overrides OPENROUTER_PROVIDER_SORT. (Phase 2)

  • order (array of strings, optional): A prioritized list of provider IDs (e.g., ["openai/gpt-4o", "anthropic/claude-3-opus"]). Overrides OPENROUTER_PROVIDER_ORDER. (Phase 2)

  • require_parameters (boolean, optional): If true, only use providers that support all specified request parameters (like tools, functions, temperature). Overrides OPENROUTER_PROVIDER_REQUIRE_PARAMETERS. (Phase 2)

  • data_collection ("allow" | "deny", optional): Specify whether providers are allowed to collect data from the request. Overrides OPENROUTER_PROVIDER_DATA_COLLECTION. (Phase 2)

  • allow_fallbacks (boolean, optional): If true (default), allows falling back to other providers if the preferred ones fail or are unavailable. If false, fails the request if preferred providers cannot be used. Overrides OPENROUTER_PROVIDER_ALLOW_FALLBACKS. (Phase 2)

Example Usage:

{
  "tool": "chat_completion",
  "arguments": {
    "model": "anthropic/claude-3-haiku",
    "messages": [
      { "role": "user", "content": "Explain the concept of quantization in AI models." }
    ],
    "max_tokens": 500,
    "provider": {
      "quantizations": ["fp16"],
      "ignore": ["openai"],
      "sort": "price",
      "order": ["anthropic/claude-3-haiku", "google/gemini-pro"],
      "require_parameters": true,
      "allow_fallbacks": false
    }
  }
}

This example requests a completion from anthropic/claude-3-haiku, limits the response to 500 tokens. It specifies provider routing options: prefer fp16 quantized models, ignore openai providers, sort remaining providers by price, prioritize anthropic/claude-3-haiku then google/gemini-pro, require the chosen provider to support all request parameters (like max_tokens), and disable fallbacks (fail if the prioritized providers cannot fulfill the request).

search_models

Search and filter available models:

interface ModelSearchRequest {
  query?: string;
  provider?: string;
  minContextLength?: number;
  capabilities?: {
    functions?: boolean;
    vision?: boolean;
  };
}

// Response: ToolResult with model list or error

get_model_info

Get detailed information about a specific model:

{
  model: string;           // Model identifier
}

validate_model

Check if a model ID is valid:

interface ModelValidationRequest {
  model: string;
}

// Response: 
// Success: { isError: false, valid: true }
// Error: { isError: true, error: "Model not found" }

Error Handling

The server provides structured errors with contextual information:

// Error response structure
{
  isError: true,
  content: [{
    type: "text",
    text: "Error: [Category] - Detailed message"
  }]
}

Common Error Categories:

• Validation Error: Invalid input parameters

• API Error: OpenRouter API communication issues

• Rate Limit: Request throttling detection

• Internal Error: Server-side processing failures

Handling Responses:

async function handleResponse(result: ToolResult) {
  if (result.isError) {
    const errorMessage = result.content[0].text;
    if (errorMessage.startsWith('Error: Rate Limit')) {
      // Handle rate limiting
    }
    // Other error handling
  } else {
    const data = JSON.parse(result.content[0].text);
    // Process successful response
  }
}

Development

See CONTRIBUTING.md for detailed information about:

• Development setup

• Project structure

• Feature implementation

• Error handling guidelines

• Tool usage examples

# Install dependencies
pnpm install

# Build project
pnpm run build

# Run tests
pnpm test

Changelog

See CHANGELOG.md for recent updates including:

• Unified response format implementation

• Enhanced error handling system

• Type-safe interface improvements

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.