Run Claude Code, Agent Team, ZeroClaw, Aider, Cline, and mini-agent on 9 free-tier providers instead of paying for the Claude API.
All free tiers — configure only the ones you have keys for.
One proxy, two API formats, nine providers — automatic fallback and rate limiting built-in.
Claude Code / Agent Team ZeroClaw / Aider / mini-agent (Anthropic API) (OpenAI API) │ │ │ POST /v1/messages │ POST /v1/chat/completions ▼ ▼ ┌─────────────────────────────────────────────────┐ │ model-router │ │ [Rate Limiter] [SHA-256 Cache] [Fallback] │ └──────────────────────┬──────────────────────────┘ │ ┌────────────────┼────────────────┐ ▼ ▼ ▼ Mistral Cerebras Groq Gemini OpenRouter Ollama ...and 3 more
Run Claude Code completely free — using your own LLM provider keys instead of the Anthropic API.
Server starts on http://localhost:8787
Any non-empty string works as API key — it's never sent upstream.
Claude Code connects and starts — your requests go to the free provider.
Claude Code's UI displays its own model labels regardless of the backend. Your requests are actually processed by whichever free provider is active (Mistral, Groq, etc.).
To verify, check the bun run dev terminal — it logs [router] streaming via mistral for every request.
Shows request counts and token usage per provider in real time.
Or ask inside Claude Code: "What model are you?" — Mistral will answer with its identity.
You hit the rate limit of your only active provider. Fix: add more provider keys to .env and set a fallback chain:
Run direct live checks against configured provider APIs to verify keys, provider accessibility, and stream functionality with minimal token usage (maxTokens: 5):
E2E tests do not run during CI or pre-commit checks. An adapter is tested only if its corresponding key exists in .env.
Works with any Anthropic or OpenAI-compatible tool.