@0xmrk/openrouter-mcp

🌐 openrouter-mcp

Talk to 500+ AI models from inside Claude Code. One file. Zero config. Just works.

"ask deepseek to review this code" — and it does. Without leaving your session.

---

🤯 Why This Exists

You're in Claude Code. You're deep in a problem. You want a second opinion from DeepSeek. Or a quick answer from Gemini. Or you want to see how GPT handles a prompt.

Before: Open a new tab. Copy context. Paste. Wait. Copy response. Paste back. Lose your flow.

After:

❯ ask deepseek to review my auth implementation

  deepseek/deepseek-r1 (142 in / 891 out · $0.0028):

  "Three issues: 1) The JWT expiry is set to 30 days which is too long..."

That's it. One sentence. The model responds inline. You never leave your session.

---

⚡ 30-Second Setup

# 1. Get an API key from https://openrouter.ai/keys

# 2. Add to Claude Code
claude mcp add openrouter -s user \
  -e OPENROUTER_API_KEY=sk-or-... \
  -- npx openrouter-mcp

# 3. Restart Claude Code. Done.

Or install from source:

git clone https://github.com/anthropics/openrouter-mcp
cd openrouter-mcp && npm install

claude mcp add openrouter -s user \
  -e OPENROUTER_API_KEY=sk-or-... \
  -- node /path/to/openrouter-mcp/index.js

---

🎯 What You Can Do

💬 Ask any model anything

❯ ask deepseek to explain this regex
❯ ask gemini to find bugs in my code
❯ ask gpt what design pattern fits here

Just say "ask [model]" and Claude routes it through OpenRouter. Token usage and cost shown on every response.

🏷️ Use aliases instead of memorizing model IDs

❯ ask the fast model to summarize this file
❯ ask the code model to optimize this function
❯ ask the reason model why this algorithm is O(n²)

🧵 Have multi-turn conversations with external models

❯ start a thread called "refactor" with deepseek — here's my database module [code]

  deepseek: "I see several issues with the connection pooling..."

❯ continue the refactor thread — what about the query builder?

  deepseek: "Building on what we discussed, the query builder should..."

The external model remembers the full conversation. Your Claude session stays clean. Delegate complex sub-tasks to a specialist model without polluting your main context.

🎭 Save personas for instant role-switching

❯ save a persona called "security_auditor" — expert at finding vulnerabilities, uses deepseek

❯ use the security_auditor persona to review my auth middleware

  deepseek (as security_auditor): "Critical: the token validation
  doesn't check for algorithm confusion attacks..."

Save a system prompt + model combo once, reuse it by name forever (within the session). Perfect for:

• code_reviewer — thorough, opinionated reviews
• translator — technical docs in any language
• explain_like_5 — simplify complex concepts
• devil_advocate — challenge your assumptions

💰 Track spending and set budgets

❯ set openrouter budget to $2

  Budget set to $2.00. Remaining: $2.0000.

❯ how much have I spent on openrouter?

  Session usage:
  - Calls: 12
  - Spent: $0.0847
  - Budget: $2.00
  - Remaining: $1.9153

Every response already shows inline cost: (142 in / 891 out · $0.0028). Set a hard budget limit and the server blocks requests when you hit it. No surprise bills.

🔍 Find the right model

❯ search openrouter models for deepseek
❯ show me free models on openrouter
❯ list google models sorted by cost
❯ what openrouter models have the largest context?

Filter by provider, search by name, sort by cost or context length. 500+ models at your fingertips.

---

🛠️ All 9 Tools

---

🏗️ Architecture

Claude Code session
  │
  │  You: "ask deepseek to review this code"
  │
  ├─ Claude sees your full conversation
  │  Composes a prompt with relevant context
  │
  ├─ MCP tool call → openrouter-mcp (stdio)
  │    │
  │    ├─ Resolves model alias/persona/default
  │    ├─ Loads thread history if applicable
  │    ├─ Checks budget limit
  │    ├─ POST openrouter.ai/api/v1/chat/completions
  │    ├─ Tracks token usage + cost
  │    └─ Returns response with metadata
  │
  └─ Claude presents the response to you

Key design decisions:

• Stateless relay — Claude handles context. The MCP server just forwards. No complex state management.
• Threads are opt-in — Only when you explicitly name a thread does the server maintain state.
• Budget is session-scoped — Resets on restart. No persistent config to manage.
• Single file — The entire server is index.js. Read it in 5 minutes.

---

⚙️ Configuration

That's it. No config files.

---

📊 What You See on Every Response

deepseek/deepseek-r1 (142 in / 891 out · $0.0028):

The actual response text here...

• Model name — which model actually responded
• Token count — prompt tokens in / completion tokens out
• Cost — estimated cost based on OpenRouter pricing
• Thread info — turn count if using a thread

---

🧩 Works With

• ✅ Claude Code (primary target)
• ✅ Any MCP-compatible client
• ✅ Node.js 18+
• ✅ macOS, Linux, Windows (WSL)

---

📦 One File

The entire server is index.js — 617 lines, 2 dependencies, zero build step. Read it, fork it, extend it.

openrouter-mcp/
├── index.js         ← the entire server
├── package.json
└── README.md

---

🤝 Contributing

PRs welcome. The bar is simple: does it make the tool more useful without making it more complex?

Good PRs:

• New model aliases as the landscape evolves
• Better cost estimation
• Smarter typo correction

Please don't:

• Add config files
• Add a build step
• Break the single-file design

---

📄 License

MIT