ClawRouter — The LLM router built for autonomous agents
A local smart router that lets AI agents use 71 LLMs without accounts or API keys, paying per request via x402 USDC micropayments.
Evidence shows: The project claims local routing and no external dependencies, but no code-level evidence for least privilege. User confirmation mechanism is absent; no user confirmation steps mentioned. Data flow transparency is partial: README describes request flow but not data storage/logging. Sensitive data handling: involves wallet private keys but no storage/encryption details. Dependency security: has overrides and CI but no vulnerability scan results. External effects: involves phone calls and payments but no user confirmation. Rollback: has upgrade tests but no rollback mechanism. Source attribution: has author and repo but publisher unverified. Deductions: lack of code-level evidence, missing user confirmation, unclear sensitive data handling.
Evidence shows: README has inconsistent numbers (e.g., free models 6 vs 71), but CI has brand number check. Dependency availability: has package.json and CI but not verified. Failure messages: tests have error handling but no user-visible failure messages. Deductions: inconsistent numbers, unclear failure messages.
Evidence shows: README details multiple usage scenarios (OpenClaw, continue.dev, Cursor, etc.) but capability boundaries unclear. Trigger precision: has routing profiles but no trigger conditions. Environment fit: has Node version requirement but no other environments. Deductions: unclear capability boundaries and trigger conditions.
Evidence shows: README is well-structured with install notes and examples, but lacks FAQ and known limitations. Naming stability: has version numbers but no naming rules. License: MIT license present. Versioning/changelog: not provided. Maintenance responsibility: has author and repo but no maintenance plan. Deductions: missing FAQ, known limitations, changelog.
Evidence shows: README claims 88% cost savings but no independent verification. Output usability: has API and CLI but no output format details. Marginal value: has unique features (x402 payments) but no comparison with alternatives. Cost-benefit: has pricing tables but no actual cost analysis. Deductions: unverified cost savings claim.
Evidence shows: README references savings-mix.json but no specific data. Cross-source corroboration: has CI and tests but no independent verification. Fact/inference separation: README mixes facts and marketing claims. Deductions: lack of verifiable data and independent verification.
- The project involves wallet private keys and payments, but no security audit or vulnerability scan evidence is provided.
- README has inconsistent numbers (e.g., free model count), which may affect user expectations.
- No user confirmation mechanism is provided, especially for operations involving phone calls and payments.
What does this agent do, and when should you use it?
ClawRouter is an open-source (MIT-licensed) smart LLM router designed for autonomous AI agents. It analyzes each request across 15 dimensions and routes to the cheapest capable model in under 1ms, entirely locally with no external API calls for routing decisions. Unlike human developers, agents cannot sign up for accounts or enter credit cards, so ClawRouter uses wallet signatures as authentication and USDC micropayments via the x402 protocol on Base and Solana. It offers 71 models from nine providers including OpenAI, Anthropic, Google, xAI, and DeepSeek, with six models completely free (NVIDIA-hosted). The tool is distributed as an npm package and can run as an OpenClaw plugin or a standalone local proxy on port 8402, compatible with any OpenAI-compatible client.
ClawRouter starts a local proxy (default port 8402) that intercepts OpenAI-compatible requests. It evaluates each request with a weighted scorer across 15 dimensions, decides the tier (simple, medium, complex, reasoning), and selects the optimal model. Users can choose routing profiles via /model auto, /model eco, /model premium, or /model free; the auto profile costs 88% less than pinning Claude Opus 5, and eco 98% less, based on a published workload mix. Payments are handled through x402: the proxy returns a 402 status with a price, the wallet signs, and the request is retried. Authentication is via a locally generated BIP-39 mnemonic deriving both EVM and Solana wallets. CLI commands include clawrouter setup (for OpenClaw integration), clawrouter doctor (AI diagnostics), and slash commands like /wallet, /exclude, and /stats. Additional features include image generation (/cr-imagegen), video generation (/videogen), image editing (/img2img), voice calls (/cr-call), and the Surf crypto data API.
- AI coding agents running in OpenClaw that need access to multiple LLMs without registration or API keys, and want to cut costs.
- Developers using OpenAI-compatible clients like continue.dev, Cursor, or VS Code who want automatic routing to the cheapest capable model instead of manual selection.
- Hobbyists and developers who want to use LLMs for free with zero balance, adding USDC later for paid models.
- Teams that handle a mix of simple and complex requests and want to avoid paying premium prices for trivial tasks.
- Users of Claude Code who want to avoid an Anthropic account and credit card, using the BRCC companion for the same routing and payment benefits.
- Agents that need to generate images, videos, or make AI voice calls directly from chat, all paid per request with USDC from the same wallet.
What are this agent's strengths and limitations?
- No accounts, API keys, or credit cards required: wallet signatures are authentication, and USDC micropayments via x402 are fully agent-friendly.
- Truly local smart routing: a 15-dimension scorer selects models in under 1ms with no external API calls, and the logic is open source.
- Significant cost savings: 88% cheaper than pinning Claude Opus 5 on auto, 98% on eco, with a reproducible methodology.
- Six free NVIDIA models with no balance or signup, making it possible to run at $0 forever.
- Works with any OpenAI-compatible client, including continue.dev, Cursor, and VS Code, and integrates as an OpenClaw plugin.
- Routing depends on BlockRun's model catalog and pricing, which is controlled upstream and might change.
- x402 payments require USDC on Base or Solana, introducing a cryptocurrency and network learning curve.
- Free models are NVIDIA-hosted, which may have latency or availability limitations in some regions.
- Standalone setup requires exact API base URL configuration (must end with
/v1/) to avoid 404 errors. - Advanced features like voice calls and image generation rely on third-party providers (Twilio, Bland.ai, BlockRun gateway), adding dependencies and complexity.
How do you install or deploy this agent?
Installing ClawRouter requires Node.js and npm. For OpenClaw, use the recommended one-shot script: curl -fsSL https://blockrun.ai/ClawRouter-update | bash and restart the gateway with openclaw gateway restart. Alternatively, install via npm globally: npm install -g @blockrun/clawrouter, then run clawrouter setup — a required step to register the plugin with OpenClaw. For standalone use, run npx @blockrun/clawrouter to start the local proxy. No signup or API key is needed; a wallet is generated on first run.
How do you use this agent?
Once running, ClawRouter acts as a local proxy on port 8402. In OpenClaw, the default model is blockrun/auto; switch profiles with /model free, /eco, /premium, etc. For continue.dev or Cursor, set the API base URL to http://localhost:8402/v1/ (ensure trailing slash), API key to x402, and model to blockrun/auto. To use paid models, send USDC to your wallet address on Base or Solana; for free use, pick any free model like free/mistral-nemotron. Monitor balances with /wallet, exclude models with /exclude add <model>, and view usage with /stats. Diagnose issues with npx @blockrun/clawrouter doctor.
How does this agent compare with similar options?
OpenRouter: ClawRouter is open source, runs locally, and uses wallet authentication and USDC payments, whereas OpenRouter requires an account, API key, and credit card, and only offers manual model selection. LiteLLM: Both are open source and local, but ClawRouter adds smart routing, wallet auth, and USDC payments; LiteLLM requires you to bring your own API keys and choose models manually. Martian: Offers smart routing but is closed-source, requires account and API key, and has no free tier. Portkey: Provides a gateway with observability but relies on subscription fees and is not agent-native.
FAQ
Is ClawRouter really free?
/model free to route across them.How do I fund my wallet for paid models?
Can I use ClawRouter without any cryptocurrency?
free/ model or use /model free. For paid models, you need USDC.Does ClawRouter work with my existing tools?
http://localhost:8402. It also integrates as an OpenClaw plugin, and BRCC serves Claude Code.What if something goes wrong?
npx @blockrun/clawrouter doctor for AI-powered diagnostics. You can also exclude problematic models with /exclude add <model>.