11 天内 916 个 Jev 开源项目，作者挑出 10 个先装

qibaz @QibazX原推发布于 2026.09.27

在 X 上打开原推

jev is 11 days old and one community catalog already counts 916 open-source projects built on it these are the 10 i'd install first. they plug into claude code, codex, postgres, a browser and a plain terminal quick context: you send jev some state plus questions with fixed answers, and it replies with probabilities > choice: pick one of your labels > score: place it on a scale you define > noul: how likely a statement is true 1. browser-use/jev-ultrafast jev picks each action and its target on the page, a small llm only types the text. google flights, Zürich to London, 7.1 seconds → http://github.com/browser-use/jev-ultrafast 2. tamaratran/fast-jev-compaction replaces the claude code compaction summary. jev scores every tool call, drops or trims the stale ones and keeps the rest word for word → http://github.com/tamaratran/fast-jev-compaction 3. jkudish/jev-mcp 11 judgment tools for any mcp agent: fact-check claims, screen pages for injected instructions, rerank results, review a diff → http://github.com/jkudish/jev-mcp 4. realZachi/pg-jev a postgres extension. where jev(tickets, 'the customer is angry') works as a normal sql filter, no embeddings or vector column → http://github.com/realZachi/pg-jev 5. uehaj/jev-semgrep grep by meaning. one japanese query finds refund requests in french, russian, german, spanish, chinese and korean → http://github.com/uehaj/jev-semgrep 6. y0usaf/pi-jev a gate for the pi coding agent. before bash, write or edit runs, jev checks for damage, leaked secrets and scope creep in about 300 ms → http://github.com/y0usaf/pi-jev 7. 0xNatoshi/jev-codex-router jev picks the model and the thinking effort for every codex call, with a safe fallback route when it errors → http://github.com/0xNatoshi/jev-codex-router 8. tamaratran/jev-pruner bash output over 10k tokens gets trimmed before claude reads it. errors, diffs and json pass through untouched → http://github.com/tamaratran/jev-pruner 9. typesafe-ai/skills the official skill from typesafe. your coding agent learns to design jev workflows from the current docs and cookbooks → http://github.com/typesafe-ai/skills 10. typesafe-ai/system-one-adapter-python the same client with gpt, claude or gemini answering instead. run your own cases through both before you switch → http://github.com/typesafe-ai/system-one-adapter-python where i'd start > new to jev: 9, then 10 > on claude code: 2 and 8 > want it offline: jaredpalmer/kev runs jev-style models from 0.8b to 27b, and the typesafe python sdk talks to it unchanged

这条是引用着另一条写的：x.com/qibazx/status/2101743128222851446