Seven Jev project ideas: browser agent, trading, Mario

Robin Delta @heyrobinaiOriginally posted 2026.09.26

Open original on X

Jev has been blowing up and I finally tried the API myself.. you don't need to understand the Jev API to use it most people overthink it -> read docs, watch videos, still confused just copy this checklist instead and skip all of that: 1. jev-ultrafast Stop making a frontier model stare at a webpage and “reason.” Jev just answers: what’s the next move, which element? A tiny model only gets called when something needs to be typed. Google Flights search lands in ~7 seconds. This is the reference build for “browser agent but not broke.” https://github.com/browser-use/jev-ultrafast 2. killmyidea Paste a startup idea. Jev scores it from a bunch of indie-hacker angles and returns KILL / FIX / SHIP. No pep talk. No founder-mode coping. Just the verdict. https://github.com/monteduro/killmyidea 3. typesafe-mario Jev plays Super Mario. No screenshots. It reads structured emulator RAM and decides: run, jump, dodge. Proof that “decision model + clean state” beats “vision model + vibes.” https://github.com/fhshaik/typesafe-mario 4. jev-trader High-frequency market making on Monad testnet. Spreads + flow in, buy/sell out. ~81ms model latency. This is the “can a judgment model keep up with a market” demo. https://github.com/jarrodwatts/jev-trader 5. Canny Coding agents love to announce “done” while tests are on fire. Canny looks at tool output, diffs, and test results, then judges whether that completion claim is real. Anti-cope layer. https://github.com/qkal/Canny 6. json-render Vercel Labs’ generative UI stack. Jev doesn’t drip JSON token-by-token. It picks the component, the props, the layout. The expensive model isn’t designing the page one comma at a time. https://github.com/vercel-labs/json-render 7. jev-drone Flight stack still owns stability and safety. Jev only does the high-level calls: climb, brake, go around the thing. Judgment layer, not a pilot replacement. https://github.com/RomanSlack/jev-drone 8. fast-jev-compaction Claude Code context is a landfill. Before every tool call, Jev scans what’s still useful, deletes the rest, and doesn’t rewrite anything. Original text stays original. Less tokens, less hallucination, less “wait why did you forget the bug.” https://github.com/tamaratran/fast-jev-compaction 9. OneVOneJev 1v1 FPS in the browser. Every tick: move, look, aim, shoot, jump. Real-time decisions where a chat model would already be dead. https://github.com/emrickgarrett/OneVOneJev 10. typesafe-mcp You just got the API key. Plug Jev into Claude Code, Claude Desktop, Codex, Pi. Choice / Score / Noul whenever you want a real decision instead of vibes-as-JSON. https://github.com/itsmostafa/typesafe-mcp 11. Blink Repo navigation without the “read the entire monorepo” tax. At each directory, Jev picks the files most relevant to the issue, then you keep digging. Less wandering, more hitting the actual file. https://github.com/ellipsis-dev/blink 12. Prism Jev doesn’t place the order. It labels the tape: toxic flow, pressure, mean reversion, etc. Your actual strategy still trades. Separation of “what’s happening” from “what we do.” https://github.com/irfndi/prism-liquidity-agent 13. agent-desktop Desktop automation that reads the accessibility tree, not pixels. Jev judges the next button, menu, or input. Computer-use energy, minus the screenshot tax. https://github.com/lahfir/agent-desktop 14. jev-curate Training data is mostly junk. Point Jev at JSONL / Parquet: quality, relevance, risk. Only the keepers go into the next training run. Filter first, train second. https://github.com/AkashPriyadarshii/jev-curate 15. jev-codex-router Don’t send every coding turn to the $$$$ model. Jev first judges how hard this round is, then picks model tier, reasoning depth, and speed mode. Spend the budget where the task actually needs it. https://github.com/0xNatoshi/jev-codex-router 16. neo4jev Drop Jev into a knowledge graph. At every node it picks the most worth-it edge, then walks. Graph search without turning every hop into an essay. https://github.com/jexp/neo4jev 17. Winnow Context garbage collection for Claude Code. Read / Bash / Grep dump a novel into the window. Jev decides which scraps actually matter for the current task. Everything else gets yeeted. https://github.com/GhalebDweikat/winnow 18. SemDecide Jev as a CLI. Classify, score, filter straight in the shell. Hook it to crawlers, CI, data pipelines. No agent ceremony. Just a gate. https://github.com/sharziki/semdecide 19. jev-review Don’t burn a senior model (or a human) on every diff. Jev pre-screens for high-risk changes, then you escalate only those. Comes with a local dashboard so you can see the triage. https://github.com/devagrawal09/jev-review 20. jev-mcp Ready-made judgment toolkit: fact-check, screen content, rank by meaning, classify, extract. The boring-but-necessary calls agents skip because a big model is too slow and too expensive to run on every claim. https://github.com/jkudish/jev-mcp