10 GitHub repos where Jev makes the decisions for Opus 5.5 agents

Asteri @Asteri_ethOriginally posted 2026.10.10

Open original on X

Opus 5.5 + Jev: 10 GitHub repos where Jev makes every call and your agents just do the work opus writes the code. Jev answers the small questions in between: which model, which chunk, which click, is it really done. code enforces the answer. 1. jev-ultrafast ▸ https://github.com/browser-use/jev-ultrafast → a browser agent from Browser Use: Jev picks the action and the exact DOM element in one round trip, a small LLM only types the text. Zürich → London on Google Flights in 7.1s 2. mu ▸ https://github.com/qybaihe/mu → a Pi-based coding agent where Jev answers at 38 decision points, like which chunks of a tool output ever reach the context. on their own test logs it cut 40% and lost 0 of 72 required lines 3. Jevonian ▸ https://github.com/xinyao27/jevonian → a local OpenAI/Anthropic-compatible proxy for Claude Code, Cursor and Codex: one Jev call picks the model and the thinking level for every turn, a local ledger logs why 4. pi-warden ▸ https://github.com/DevMortimer/pi-warden → watches your agent for rule breaks, risky actions, stuck loops and fake "done" claims. 1,168 real sessions: 419 risky actions held before they ran, 76% of fake "done"s turned into real test runs 5. Oko ▸ https://github.com/bartlomein/oko → Jev ranks which code snippets actually answer the agent's question, Oko serves only those over MCP. 60-63% fewer tokens in their Codex, Claude Code and OpenCode runs 6. jev-pref ▸ https://github.com/doeixd/jev-pref → write your project preferences once in jev-pref.json, Jev checks every diff against them and feeds the findings back to the agent. setup is one line: npx jev-pref setup 7. taste-lint ▸ https://github.com/mblode/taste-lint → catches AI slop in your UI, copy and agent instructions before it ships. 28 hard checks block the commit, the judgment calls go to Jev and come back with a probability 8. jev-review ▸ https://github.com/devagrawal09/jev-review → a staged code review with a local dashboard: Jev picks the risky files, the evidence, the mechanism and the severity, code applies the thresholds 9. opencode-jev-orchestrator ▸ https://github.com/aaronshaf/opencode-jev-orchestrator → the parent model stays fixed so the cache stays warm, Jev spots the hard turns and hands them to temporary subagents 10. yoshi ▸ https://github.com/compozy/yoshi → a Claude Code/Codex proxy where Jev decides which old tool output can leave the request, tool IDs and structure stay byte-identical. still a proof of concept and slower than baseline for now same split in all 10: > Opus does the work > Jev makes the calls > Code enforces them Fork any of these Jev builds, then read the full Jev setup below ↓

This post quotes another post: x.com/asteri_eth/status/2104901781725598103