Weave's open-source router auto-picks the best LLM per request for Claude, Codex, Cursor

Engineering-intelligence company Weave has released Router, a drop-in proxy that sits between coding tools and the major model providers and chooses which model to use on a per-request basis. Rather than relying on prompt-based heuristics, it uses a small on-device embedding model and a cluster scorer derived from the Avengers-Pro routing research to score each turn and dispatch it to the strongest available model. The proxy speaks the Anthropic Messages, OpenAI Chat Completions, and Gemini native APIs — with streaming, tool use, and vision — and can also reach open-weight models like DeepSeek, Kimi, GLM, Qwen, Llama, and Mistral through OpenRouter or any OpenAI-compatible endpoint.

Setup is aimed at low friction: an npx @workweave/router installer wires up Claude Code, OpenAI’s Codex CLI, or opencode against a hosted router, patching each tool’s config and leaving the rest untouched, with per-client on/off toggles and slash commands. Cursor is supported in an early beta via a base-URL override. Teams that prefer to self-host can run the router and its dashboard locally on port 8080 backed by Postgres. Provider keys stay on the user’s machine, encrypted at rest (bring-your-own-key), and are kept distinct from the rk_ router key that clients pass as a bearer token.

The pitch is observability and cost-efficiency without lock-in: OTLP traces ship by default and can flow into Honeycomb, Datadog, Grafana, or Weave’s own dashboard. The practical appeal is routing cheap requests to cheap models and hard ones to frontier models automatically, but the trade-offs worth watching are the accuracy of the embedder’s routing decisions, the added proxy hop in the request path, and trusting a third-party layer with traffic between developer tools and model providers.