Let your coding agent pick a cheaper or stronger AI model per step
Made by Weave
View profile on GitHub (opens in a new tab)Weave Router sits between a coding agent (Claude Code, Codex, opencode, pi or Cursor) and the AI providers, and picks which model answers each request: a strong model for hard debugging, a cheaper one for simple edits.
It is a drop-in proxy (a stand-in server that forwards requests) speaking the Anthropic, OpenAI and Gemini formats, and the open-source version scores each request with a small embedding model on your own machine. The company behind it, Workweave, reports that its newest routing model matched GPT-6 Astra on two coding benchmarks at about half the cost.
Can I use this?
You'll need Node.js 18+ and a coding agent for the hosted quick start; Postgres and provider keys to self-host
Worth knowing Made by Workweave, a company that also sells a hosted version, and the one-command quick start routes your agent through that hosted service. On Hacker News the creator said the trained routing model behind the headline benchmark is not released as open weights; the open-source router defaults to a simpler scorer, and an older frozen routing model can be self-hosted with a Google API key. The benchmark figures are the company's own. It is Apache-2.0 from release router-v0.2.24; earlier releases kept their original licenses. BuildTube has not run or verified this software.
Why it's here
- 5,572 people have starred it on GitHub.
- Shown on Hacker News (Show HN) on 30 September, with 122 points.
Numbers as of 6 October 2026, from BuildTube’s latest daily refresh.
Behind it
See the code on GitHubBuildTube has not run or verified this project. Everything above is written from what the creator published.
Made this?