Orcarouter MCP Server

What Continuum-AI-Corp/orcarouter-mcp-server MCP server does

Continuum-AI-Corp/orcarouter-mcp-server MCP server connects MCP clients to OrcaRouter for two related tasks: exploring the available LLM catalog and sending chat completion requests. The catalog tools expose provider names, model metadata, pricing, context limits, capabilities, supported endpoints, latency percentiles, and release dates. These read-only operations do not require an OrcaRouter API key.

The chat tool sends a single-turn request and returns the assistant’s response text. It can use the workspace’s automatic router by default, address another OrcaRouter router, or call a provider/model directly. This makes the server suitable for agents that need to compare models before choosing one, or that should delegate model selection to OrcaRouter.

How it works

The package runs as a local Node.js MCP process and communicates with the client through the MCP protocol. MCP clients discover the available input schemas through tools/list, so the client can present or invoke the four exposed tools without a separate tool-definition setup.

orcarouter_providers_list returns provider identifiers and model counts. Those identifiers can then be supplied to orcarouter_models_list, whose provider, capability, and minimum-context filters are combined server-side. orcarouter_model_card is intended for a known model ID and returns a fuller record than the catalog listing.

For chat, the primary model is attempted first. An optional models array can add up to four fallbacks, for a maximum chain of five models including the primary. Requests aimed at OpenAI reasoning models have max_tokens translated to max_completion_tokens at the wire level. Failures, including missing credentials, rate limits, and upstream outages, are returned as error text with isError: true.

Setup and configuration

Install and run Continuum-AI-Corp/orcarouter-mcp-server MCP server with Node.js 18 or newer:

npx -y @orcarouter/mcp

Add the command to an MCP client configuration, then provide ORCAROUTER_API_KEY for chat calls. The README documents configurations for Claude Desktop, Claude Code, Cursor, and Windsurf; the server also works with other clients that implement MCP. The key is read from the environment, is not logged, and is sent only to the OrcaRouter API.

Configuration variables include:

• ORCAROUTER_API_KEY: required only by orcarouter_chat.
• ORCAROUTER_BASE_URL: optional API base URL; the default is https://api.orcarouter.ai.
• ORCAROUTER_REQUEST_TIMEOUT: optional per-request timeout in seconds; the default is 300.

Tools and capabilities

Continuum-AI-Corp/orcarouter-mcp-server MCP server exposes these tools:

• orcarouter_chat: Send a single-turn chat request with router, direct model, and fallback-chain options.
• orcarouter_models_list: Browse the complete catalog or filter it by provider, capability, and minimum context length.
• orcarouter_model_card: Retrieve detailed metadata for one model, including pricing, modalities, context window, output limit, endpoints, latency, and release date.
• orcarouter_providers_list: List providers with IDs, display names, icon URLs, and model counts.

The catalog can be used before authentication to compare available models and pricing. Chat routing is the only listed capability that needs the OrcaRouter credential.

Limitations and notes

The server performs single-turn chat requests rather than describing a multi-turn conversation or agent memory feature. Catalog results are read-only. Provider and model identifiers must match the OrcaRouter catalog; the provider-list tool is the documented way to discover valid provider IDs. The fallback chain is limited to five total entries, including the primary model. API availability, rate limits, and provider outages can still cause chat errors even when the MCP process is running.