OpenRouter MCP — One Server for 300+ LLMs in Claude Code

Quick Use

1. Sign up at openrouter.ai → API key (starts sk-or-v1-)
2. Add the JSON snippet below to your MCP config
3. Restart your MCP host — openrouter_chat and friends are now available

Intro

OpenRouter MCP wraps OpenRouter's universal LLM gateway as a Model Context Protocol server. Inside Claude Code, Cursor, or Codex CLI, you can call openrouter_chat and pick any of 300+ models per task — switch from Claude to GPT-4o to Llama 3.3 mid-session without changing your host. Best for: agents that want to delegate sub-tasks to cheaper or specialized models. Works with: any MCP host. Setup time: 3 minutes.

MCP config

{
  "mcpServers": {
    "openrouter": {
      "command": "npx",
      "args": ["-y", "openrouter-mcp"],
      "env": {
        "OPENROUTER_API_KEY": "sk-or-v1-...",
        "OPENROUTER_DEFAULT_MODEL": "anthropic/claude-3.5-haiku"
      }
    }
  }
}

Tools exposed

Common patterns

> use the openrouter_chat tool with model "deepseek/deepseek-chat" to summarize
  this 50-page PDF (cheap model for bulk text)

> use openrouter_list_models to find models with <$0.50/1M input cost and
  >100K context window

> use openrouter_chat with model "openai/o1" for the hardest reasoning step,
  then back to claude-haiku for the boilerplate

Why use OpenRouter MCP vs direct provider MCPs

A separate MCP per provider (Anthropic MCP, OpenAI MCP, …) means N config blocks, N keys, N tool prefixes. OpenRouter MCP gives you one config / one key / one tool prefix for all of them. Trade-off: you depend on OpenRouter's availability for everything (their uptime is good but it's a single point of failure).

FAQ

Q: Is OpenRouter MCP official? A: OpenRouter has multiple community-built MCP wrappers (search npm for 'openrouter-mcp'). They all use OpenRouter's REST API. The most active is openrouter-mcp by the community. OpenRouter itself doesn't ship an official one yet.

Q: Can I cap spend per MCP call? A: Yes — set max_tokens in each call, and use OpenRouter's account-level spend limit. Combined with PostHog observability you can also enforce per-feature budgets.

Q: Does this work for non-coding agents? A: Yes — OpenRouter's tools work for any text task. The MCP just exposes them to MCP hosts. Same patterns work for chat apps, content generation, data extraction agents.

Source & Thanks

Built by community / OpenRouter. MIT-licensed wrappers.

openrouter.ai — OpenRouter platform