Jev community projects: browser agents, code review, skill routing
A roundup of community projects built on Jev, including browser-use's ultrafast flights demo, an MCP tool, stepwise PR review, and skill routing at 94.4% accuracy, treating Jev as an AI if statement.
Jev @CompleteSkeptic dropped a few days ago, and the most interesting part isn’t the benchmark. It’s what people are already building with it. Browser agents. Code review. Skill routing. NPCs. Trading bots. Drone logic. Email triage. The pattern is becoming pretty clear: Jev is turning into an “AI if statement.” A thread on the best community examples I’ve found ↓ ⸻ 1/ Browser agents @browser_use built jev-ultrafast. Their Flights demo reportedly finishes in ~7.1s. Makes sense when you look at what a browser agent actually does all day: click this or that? continue or retry? did the page satisfy the condition? which action should happen next? A lot of that doesn’t need full text generation. It just needs a fast decision. http://github.com/browser-use/jev-ultrafast ⸻ 2/ Jev as an MCP tool @jkudish built jev-mcp. It lets agents use Jev for things like: claim verification prompt injection checks ranking classification This is probably one of the most obvious fits. Instead of: LLM → reasoning → JSON → parse → if you can move toward: Jev → probabilities → if http://github.com/jkudish/jev-mcp ⸻ 3/ Code review @devagrawal09 built jev-review. Instead of asking one model: “review this entire PR” the workflow breaks review into smaller decisions: is there a likely bug? did the API behavior change? is there a security concern? does this need human review? That decomposition feels very native to this kind of model. http://github.com/devagrawal09/jev-review ⸻ 4/ Skill routing GitHub user GodsBoy built jev-agent-skill-router. The author reports 94.4% routing accuracy. This one is especially interesting to me. Once agents have 50, 100, or 500 skills, deciding which skill to call becomes its own model problem. Do we really want a giant LLM rereading every tool description on every turn? Probably not. http://github.com/GodsBoy/jev-agent-skill-router ⸻ 5/ NPC behavior @haiderTheDev built Hollow Creek. Instead of using a large model for every NPC action, Jev can sit in the loop and answer smaller questions: scared? approach? run? attack? trust the player? That feels much closer to a game “reflex system” than a chatbot. http://hollow-creek-sigma.vercel.app ⸻ 6/ Trading every block @jarrodwatts built jev-trader. Every block, Jev decides: BUY / SELL / HOLD No idea if it makes money 😅 But the interesting part is the frequency. There are lots of places where nobody would previously call an LLM because the decision is too small, too frequent, or too latency-sensitive. That changes if inference becomes cheap enough. http://github.com/jarrodwatts/jev-trader ⸻ 7/ Drone decision layer GitHub user RomanSlack built jev-drone. The interesting architecture isn’t “AI controls the whole drone.” It’s more like: large model = planning fast decision model = tactical reflexes That separation may end up being important far beyond drones. http://github.com/RomanSlack/jev-drone ⸻ 8/ One score per pixel GitHub user rivianpratama built JevPixelArt. Each pixel effectively becomes a scoring decision. This is exactly the kind of thing that sounds absurd with a normal LLM. But once inference is cheap enough, suddenly people start putting models inside loops that used to be completely deterministic. Demo: http://jevpixel.vercel.app Repo: http://github.com/rivianpratama/JevPixelArt ⸻ 9/ The boring business use cases may matter more Near Here tested Jev for local event validation. They reported 96% on 50 examples. ActionBox explored refund review with human escalation. @DisruptConcept used it for desktop email triage. There’s also an anonymous legal-tech test on 9,840 Enron emails. None of these are flashy. But classification, filtering, routing, moderation, prioritization and review are basically everywhere inside real software. A huge amount of business logic is just: fuzzy if statements.