An LLM router that signs transactions, not NDAs
ClawRouter lets autonomous agents pay for AI with USDC micropayments instead of credit cards, because agents can't fill out Stripe forms.

What it does ClawRouter is a local TypeScript proxy that sits between your AI client and 55+ LLM providers. It scores each request across 15 dimensions, picks the cheapest capable model, and routes it in under 1ms. Payment happens per-request via USDC on Base or Solana using the x402 protocol — no accounts, no API keys, just wallet signatures. Ten NVIDIA models are free without any crypto setup.
The interesting bit The “agent-native” framing is the real angle. Every other router assumes a human with a credit card and a dashboard login. ClawRouter assumes the opposite: the caller is code that can sign a blockchain transaction but can’t click “I agree to terms.” The x402 micropayment protocol turns HTTP 402 Payment Required into a machine-payable endpoint.
Key highlights
- 100% local routing, zero external API calls for the routing decision itself
- Wallet signature replaces API key authentication
- Four routing profiles:
free,auto(default),eco,premium - Standalone proxy mode for any OpenAI-compatible client (continue.dev, Cursor, etc.)
- OpenClaw plugin with one-shot install script
- Also supports image generation via
/cr-imagegenwith multiple providers
Caveats
- The “10 free models” and “8 free models” claims differ between sections of the README; the exact free count is unclear
- Bare
npm install -gleaves OpenClaw integration broken unless you runclawrouter setupafterward - The comparison table claims “55+ models” but the routing examples show model names that don’t map cleanly to real-world versions (e.g.,
claude-opus-4.8,gpt-5.3-codex) — verify actual model availability against live endpoints
Verdict Worth a look if you’re building autonomous agents that need to pay their own way, or if you want smart routing without cloud dependencies. Skip it if you’re a human who already has API keys and a working credit card — OpenRouter or LiteLLM may be simpler.