HydraMCP

An MCP server that lets Claude Code query any LLM through your existing subscriptions. No extra API keys, no per-token billing. Just your monthly subscriptions working together from one terminal.

What It Looks Like

Four models, four ecosystems, one prompt. This is real output from a live session:

All four independently found the same async bug. Then each one caught something different the others missed. GPT-5 was fastest, Gemini was most thorough, Claude gave the fix direction, Qwen explained the why. Different training, different strengths, one comparison.

And this is consensus with a local judge:

Three cloud models polled, local model judging them. 686ms to evaluate agreement, no quota used.

Five Tools

• list_models - See what's available across all your providers

• ask_model - Query any model and get a response back

• compare_models - Same prompt to 2-5 models in parallel, side by side with brief/detailed format

• consensus - Poll 3-7 models, a judge model evaluates agreement, returns one answer with a confidence score

• synthesize - Fan out to multiple models, then combine their best ideas into one answer that's better than any individual response

From inside Claude Code you just say things like:

• "ask gpt-5 to review this function"

• "compare gemini and claude on this approach"

• "get consensus from 3 models on whether this is thread safe"

• "synthesize responses from gpt-5, gemini, claude, and qwen on how to design this API"

It just works. No browser tabs, no copy pasting between apps.

How It Works

HydraMCP sits between Claude Code and your model providers. It routes requests to the right backend, runs comparisons in parallel, and formats results to keep your context window small.

The consensus tool uses an LLM-as-judge approach. Instead of naive keyword matching, it picks a model not in the poll and has it evaluate whether the responses actually agree. It understands that "start with a monolith" and "monolith because it's simpler" are the same answer.

The synthesize tool goes further. It collects responses from multiple models, then a synthesizer model reads all of them and builds one combined answer. Best structure from one, best insights from another, best examples from a third. The result is better than any single model could produce alone.

Setup

You need Node.js 18+, Claude Code, and at least one backend. The whole process takes about 5 minutes.

Step 1: Set Up a Backend

You need at least one of these. Both is ideal.

CLIProxyAPI (Cloud Models)

CLIProxyAPI turns your existing subscriptions (ChatGPT Plus, Claude Pro, Gemini, etc.) into a local API. You authenticate once per provider, and it handles the rest.

Get the binary:

Create

The API key is a local passphrase you make up. It sits between HydraMCP and CLIProxyAPI on your machine. Not a provider key.

Authenticate your subscriptions:

Each subscription is one login command. A browser opens, you sign in with your existing account, done.

You can authenticate as many as you want. Each one adds models to your pool. Run the same command again with a different account to add multiple credentials per provider.

Start it:

Runs on localhost:8317. Leave it running.

Ollama (Local Models)

Install Ollama, then pull a model:

Runs on localhost:11434 by default. Good for fast operations like judging consensus without using cloud quota.

Step 2: Install HydraMCP

Step 3: Configure

Edit .env to match your backends:

The CLIPROXYAPI_KEY should match the key you put in config.yaml.

Step 4: Register with Claude Code

Restart Claude Code. Type "list models" and you should see everything you authenticated.

Model Routing

You can target specific backends with prefixes:

• cliproxy/gpt-5 - explicitly use CLIProxyAPI

• ollama/qwen2.5-coder:14b - explicitly use Ollama

• gpt-5 - auto-detect (tries each provider until one handles it)

Credits

• CLIProxyAPI powers the subscription-based cloud backend

• Ollama powers the local model backend

• Built with the MCP SDK and Zod

I built the MCP tool layer, routing logic, and multi-model orchestration on top of these. Credit where it's due.

Contributing

Want to add a provider? The interface is simple. Check src/providers/provider.ts for the contract and src/providers/ollama.ts for a working example. Implement healthCheck(), listModels(), and query(), register it in src/index.ts, and you're done.

Providers we'd love to see:

• LM Studio

• OpenRouter

• Direct API keys (OpenAI, Anthropic, Google)

• Anything else that speaks HTTP

License

MIT