Awesome Jev

September 21, 2026 · View on GitHub

Awesome PRs Welcome

A curated list of resources for Jev — TypeSafe AI's "System One" model that returns typed, calibrated decisions instead of text.

Jev doesn't chat. You send it structured state; it returns a probability distribution over your predefined options in 70–500ms, at $0.042 per million input tokens with output free. It was built by Diogo Almeida (ex-OpenAI, co-inventor of RLHF/ChatGPT), emerged from stealth in September 2026 with $40M in funding, and is currently in waitlist-only early access.

This list is maintained by RoboKrunch — we benchmark Chinese edge-AI hardware and run our own real-measurement Jev experiments (300 real API calls, real latency, real bills). Entries with our own measured numbers are marked 📊.

No official paper. TypeSafe has not published a formal paper on Jev as of 2026-09-20. If one appears, it goes here first.

Contents

Official

Models & API

  • Native API: POST https://api.typesafe.ai/v1/systemone — structured DOM/state in, {operation, target} style typed decisions out. One call can carry many questions in parallel.
  • Question primitives: choice (pick from up to 255 options, per-option probabilities + confidence), noul (yes/no as a single probability), score (numerical).
  • Pricing: $0.042 / 1M input tokens, $0 output. No tokens wasted on explanations.
  • Latency envelope: TypeSafe quotes 70–500ms; independent testers measure p50 ~0.2–0.8s depending on task and route (see Benchmarks).

Projects & Code

Browser agents, mobile agents, guardrails — the fastest-moving category. Starred entries have verified repos.

  • browser-use/jev-ultrafast — Browser agent with a dynamic indexed action space. Jev picks the operation + element; a small LLM writes text only for TYPE_TEXT. Zürich → London on Google Flights in ~7.1s.
  • droidrun/mobile-jev — Mobile agent for Mobilerun: Jev drives a real Android phone, opens Uber, enters SFO → Golden Gate Bridge route, reaches payment selection. 9 actions in ~21s (no booking completed).
  • ClemensSchartmueller/jev-guard — High-speed safety gate for Claude Code, Codex CLI, and Antigravity. Intercepts tool calls, does sub-ms local checks, then asks Jev to score blast radius / reversibility / destructive potential.
  • jkudish/jev-browser — Browser automation with Jev.
  • jkudish/jev-mcp — MCP server for Jev decisions.
  • SadiqOnGithub/jev-lab — Experiments playground.
  • tumf/jev-cli — CLI for Jev.
  • steventsao/jevzen — Jev experiments.
  • zhuyansen/jev-search-rerank-eval — Search rerank evaluation with Jev.
  • nekuda-ai/WindTunnel — Agent testing harness with Jev.
  • Laya (ConvAI Innovations) — Apache-2.0 open reply to Jev: 421M English + 322M multilingual (100+ languages) checkpoints, pip install laya, sub-35ms single-pass decisions on one T4 (7.2ms batched), zero API bill. Vendor reports 0.766 vs Jev's published 0.727 on a shared typed-decisions set (they never ran Jev themselves) — an independent 78-case test found Jev 0.974 vs Laya 0.590, Laya fastest at 30ms/case vs ~302ms for Jev (small sample).
  • pngwn/system-one-qwen3.5-4b-scorer-v2b — Community single-pass scorer on Qwen3.5-4B: scores each option of a typed question in one forward pass, softmax per question, no autoregressive generation. Val split: 0.803 accuracy, ECE 0.022. Non-commercial (training data is CC-BY-NC-4.0).
  • askjev.ai — "Ask Jev anything. It won't answer. It will judge." Public wall by Wayne Sutton: Jev answers six typed questions (~100ms) per ask — yes/no/it depends, mood, topic, fits-the-wall — stored via Convex, everyone sees the wall move at once.
  • r-ms/mini-jev — "Fully local mini Jev": a Jev-style typed-decision interface on a frozen Qwen3-4B — read the option letter's logits instead of generating JSON. Preregistered experiment with a teaching bench; not production-ready.
  • robipop22/Jev-is-odd — TypeScript micro-benchmark: ask Jev whether a number is odd, logging real token usage and client-side latency. 20/20 fixed cases, median/p95 round-trip 294.3/364.8ms (network included), whole run $0.00029 (jev-1.13.0, measured 2026-09-18).
  • moritzkremb/jev-voice-browser — Voice-controlled real browser: Jev decides intent + target in ~300ms per spoken word (p50 ~300ms, 3–6k input tokens/request), Playwright acts — often before you finish the sentence. 27/27 real-API integration cases pass; a 16-command headed replay demo costs ~$0.01 (measured 2026-09-17).

Spotted in the wild (seen on X / in roundups, repo link wanted — PRs welcome):

  • jevmeter — scores every sentence of a debate for ~$0.05.
  • Typewriter (Steve Krouse) — updates 16 judgments live as you type.
  • heist-one — Jev runs the guards in a heist game; also completed StarCraft's first combat mission.
  • pg-jev — plain-language filters for Postgres.
  • HA-Jev — turns Jev answers into Home Assistant entities.
  • sift — Chrome extension labeling every X timeline post (substance, humour, promo, junk, AI-written) with Jev decisions.
  • jev-block-android-ad — Android notification/SMS noise gate; fails open on OTP codes.
  • jev-cvss — CVE description → CVSS v3.1 metric predictions.

Benchmarks & Evaluations

Numbers with sources. Vendor claims are labeled as such.

  • 📊 RoboKrunch: 10,000-robot fleet triage — 300 real decisions API calls against a simulated warehouse AMR fleet (bilingual CN/EN incidents), 3 simultaneous judgments per call (human escalation, owning team, urgency 0–2). 300/300 succeeded, p50 0.53s, p95 0.81s, $0.0000246/decision ($24.57 per million). Fleet-scale model (10K robots × 48 decisions/day × 30d): $354/mo vs $1,814/mo for GPT-4o-mini — 5.1×. Caveats: incidents are simulated (template agreement 91.3%, not production accuracy); GPT-4o-mini cost is estimated, not measured.
  • 📊 RoboKrunch: Jev vs self-hosted ModernBERT — ModernBERT-base (149M params) on a 2-core CPU: p50 169ms vs Jev's 527ms (~3× faster) but one label vs three judgments. Infra crossover ≈ 977K decisions/month (~678 robots at 48/day) before self-hosting beats Jev on cost. Conclusion: Jev's edge is zero training, zero labeling, zero ops — not raw speed.
  • Aman Kumar: "Testing Jev on public and private data — classifier or filter?" — ~16,000 real calls over two days. On short-text classification, level with or ahead of gpt-5.4-mini and gpt-5.6-luna on 3 of 4 public sets, at 5–56× lower cost, median under a second. Probabilities hold near 0 and 1 (confident answers right 90–100%); the middle is a coin flip. Verdict: "a filter, not a replacement" — Jev takes the calls it's confident about, the rest fall back to your existing model.
  • TypeSafe's own workflow evals (via MarkTechPost) — Jev 0.114s / $0.000081 vs GPT-5.6 Terra 8.566s / $0.013880 (claimed 193.6× faster, 444.6× cheaper). Fine print: reference answer is the average of GPT-6 Astra and Fable 5.1; TypeSafe's own team wrote the workflows; TypeSafe says these gains sit at the high end and it cannot prove the price is unsubsidized.
  • Vercel: command safety — Guillermo Rauch reported Jev up to 18× faster at p95 than GPT Luna for command-safety review, and more accurate (via engineer Pranit Sharma's benchmark).
  • Bryo AI: email triage — CTO Nikhil Mudholkar found Gemini slightly more accurate but 10–20× more expensive than Jev.
  • @ebrain.lab: Jev vs Claude Opus on 40 Korean sentences — One API call per sentence: Jev 40/40 correct in 1.9s for $0.0012; Claude Opus also 40/40 but ~24× more expensive. Caveat: prompting Jev like an LLM with the whole document dropped accuracy to 62% — commenters note batching increases drift.
  • @simplifyinai: 27 questions, one parallel request — Jev answered 27 typed questions in a single parallel request in 0.11s for $0.000083, head-to-head vs GPT-5.6-texta on the same task (community test).
  • anisselbd/jev-phishing-bench — Jev (jev-1.13.0) vs Claude Haiku 4.5 on 2,000 phishing emails (results 2026-09-17; reproducible repo with calibration audit). Jev: 62.6% accuracy, 43.2% recall on phishing, ECE 0.154, p50 239ms, $0.038 per 1K emails. Haiku 4.5: 81.3% accuracy, ECE 0.097, p50 687ms, $0.462 per 1K. Verdict: Jev loses on accuracy, wins on speed and cost.
  • OpenRouter: Jev vs 4 LLMs on a 30-class task (via TipRanks) — OpenRouter's own benchmark, 200 synthetic cases: Jev >5× faster than the next fastest model, matched top LLMs on accuracy, second cheapest behind Qwen3.8 Flash.
  • Matthew O'Riordan's Pong latency showdown (via RuntimeWire) — Four lanes, same game state, recorded 2026-09-17 via Vercel AI Gateway: Jev averaged 227ms/decision (p95 400ms); Gemini 3.8 Flash 3.2s, Claude Haiku 4.5 2.5s, GPT-5.6 Sol 3.5s. In the first 12s Jev returned 47 decisions vs 3/2/2 for the chat models. Caveat: a latency demo, not a strategy test — the chat models picked the correct move 95–100% of the time, and every model got structured state (no vision).
  • FazalAAli/jev-robotics-demo — Jev vs Claude Opus 5 driving a simulated Franka arm + Allegro hand in MuJoCo (stack a blue cube on a red one): one recorded run each — Jev 19.1s / $0.0006 vs Opus 158.8s / $0.75 (clean first-try Opus runs took ~55s / $0.19). Design: code owns the physics, Jev owns the judgment.

Demos on X

Single-post links rot fast; handles + what they showed. PRs with direct links welcome.

  • @gregpr07 — Browser Use + Jev, Google Flights in ~7s.
  • @CompleteSkeptic — Jev plays Doom.
  • @faadilhshaik — Super Mario Bros with Jev.
  • @nutlope — 1,018 AI papers classified for $0.08.
  • @iannuttall — 3,282 X posts, 4.25M tokens, $0.1282.
  • @TheMattBerman — 724 ads analyzed in 40s for $0.09.
  • @rileybrown — 500 emails triaged for $0.035.
  • @tdinh_me — YouTube sponsor-skip browser extension.
  • @0xidanlevin — WebMCP benchmark with Jev.
  • @Steve8708 — "Jev is awesome but for the love of god please STOP posting fake demos" — the thread that forced the fake-demo vs real-measurement debate (148K views).
  • @marckohlbrugge — Voice computer assistant with Jev: local Whisper listens, Jev classifies spoken intent against a live accessibility tree fed by a small Swift app, all real-time (early build; via the madewithjev.com showcase).

Videos

Tutorials & Guides

Hands-on, not hot takes. Entries here were checked for real code or real runs.

  • Avi Chawla: "Build your own Jev (100% local)" (X article, 2026-09-20) — Recreates Jev's fixed-answer scoring on a local open LLM via SGLang's /v1/score: read the option letter's logits, softmax over the labels, zero text generation. Real run on Qwen2.5-0.5B-Instruct (logits 25.28/24.50/21.19 → 0.68/0.31/0.01), plus a 100-case benchmark app pitting the scoring lane against a generation lane. ~5,000 words, 107K views.
  • Movez: "Jev Engineering: how to build the fastest AI Agent Brain in 10 Steps" (X article, 2026-09-19) — Ten-step guide to wiring Jev in as an agent's decision brain: Playground, SDK, a chief.py decision router, dynamic menus, parallel questions, guardrails, cost math, five production use cases. Caveat: most headline numbers (flights in 7s for $0.0039, 1,018 papers for $0.08) are quoted community experiments, and "200× faster, 400× cheaper" is the author's claim — not an independent measurement. 302K views.

News & Articles

Papers

None. TypeSafe has not published a formal paper on Jev or RLCD as of 2026-09-20. This section will not be padded with loosely-related arXiv links — when the paper drops, it goes here.

Community

Contributing

PRs welcome. Rules:

  1. One line of description per entry, with a real link. No link, no merge (post it in the issue first and we'll find it).
  2. Numbers need sources. "10× faster" with no methodology gets cut.
  3. No affiliate links, no waitlist-farming, no fake demos — the @Steve8708 rule: if the demo isn't real, it doesn't belong here.

License

This list is released under CC0 1.0 — do whatever you want with it.

Maintained by RoboKrunch — we benchmark Chinese edge-AI hardware and measure what AI actually costs in the physical world. Our Jev experiments live at robokrunch/jev-physical-ai.