Project
Jev Ultrafast (Browser Use)
Browser Use's fast browser agent where Jev picks the operation and the DOM element in a single request; a small LLM is only invoked to write text when typing is required. Demoed on a Google Flights search.
Projects built with System One AI models
Software people have built on top of a System One model, nearly all of it on Jev and nearly all of it written in the days after launch. Some are toys meant to show a model deciding something inside a loop that is already running; others are ordinary products where a typed answer replaced a parsed language-model response. Read the commit dates before you judge any of it.
89 projects
Showing 89 of 89
All entries
- Project
openjev
Attempt to run a Jev-style System One decision model locally on a single RTX 3090. Companion site at openjev.com.
- Project
jevlike
Open reimplementation that trains a small model to choose among a changing list of text options, emitting one probability per option in a single forward pass. Ships Doom, chess and Wikispeedia demos.
Choice - Project
jev-trader
Makes one AI trade decision every Monad block on the Kuru MON-USDC market using Jev. Live at jev-trader.vercel.app.
- Project
Ran Jev against an existing classifier eval that previously used Gemini 2.5 Flash Lite
Vercel CTO reports Jev saturated an existing classifier eval that had used Gemini 2.5 Flash Lite and ran about 6x faster. Original post; TypeSafe's quote-tweet is a separate item.
- Project
typesafe-mario
Agent that plays Super Mario Bros. by feeding structured emulator state to Jev and acting on the returned controller action.
ChoiceReal-time control - Project
jev-review
Staged code-review workflow with a local dashboard, where Jev judges code changes or a whole codebase.
- Project
typesafe-computer-use
macOS computer-use loop that OCRs the screen, classifies the next action with Jev, then clicks. Author reports roughly $0.0002 per step.
- Project
usenotra/notra
Generative engine optimization platform that tracks brand mentions across ChatGPT, Claude, Gemini and Perplexity. Its NOTRA_JEV_CLASSIFIERS setting moves the chat router, the GEO judge and the feedback classifier off their LLMs and onto typesafe-ai/jev, and routes the eve agent's model choice the same way.
- Project
kyotofin/tax-doc-classifier
A TypeScript classifier that sends each PDF page's text to Jev as a choice over 261 IRS forms and 7 page kinds, with a second call only for five corporate forms and their schedules. Its eval reports 0 wrong pages on 314 filled TaxCalcBench forms at about $0.001 per page, 34x cheaper and 6x faster than the Sonnet pipeline it replaced.
- Project
jev-eval-agent
Personal-assistant agent built with Vercel's eve framework and 100 mocked tools, served through OpenRouter. Measures how many steps the agent needs when the LLM picks the tool itself versus when Jev picks it via a 101-option Choice.
- Project
jev-drone
Camera-only autonomous drone simulated in MuJoCo with Jev in the control loop at 2.5 Hz.
- Project
mobile-jev
Standalone Android agent for Mobilerun where Jev makes every action decision, with a React studio UI and an Uber booking demo.
- Project
unclutter
WXT browser extension that removes page clutter using Jev decisions, with reusable template rules per site.
- Project
ChetasLua/jevmeter
Puts a live Jev meter on any video: every sentence is scored, every dodge flagged, and the result rendered as a 16:9 edit. The repo reports 99% preset accuracy on a held-out set and about $0.05 for a full debate.
- Project
realZachi/typesafe-adblock
Chrome extension that finds ad-shaped DOM elements in plain code, then asks Jev whether each candidate is an ad and pops it off the page. Bring your own key and there is no backend, and the author calls it a fun side project rather than a real ad blocker.
- Project
typesafe-ai-benchmark
Runs Qwen 3.8 27B on Cerebras and Jev side by side across seven synthetic workloads, plus a local Needle 3 evaluation on the same contracts. Records validated outputs, mistakes, latency, token usage and estimated cost.
- Project
Jev-Moderation-Bot
Real-time Discord moderation bot that runs Jev evaluations over message text and metadata in parallel to catch phishing, spam and social engineering, with a progressive escalation ladder.
- Project
superagents-lab/jev-search
Web search in plain language where Jev chooses sources, time ranges and search terms, then scores each result for relevance, with retrieval through Search1API. Returns links and snippets with visible relevance scores and editable filters, and writes no answers.
- Project
blink
Codebase search that finds files matching a natural-language request through repeated Jev-guided passes over a directory tree.
- Project
jexp/neo4jev
Navigates a Neo4j graph one hop at a time: the outgoing relationships of each node become Choice options and Jev returns a distribution over which to follow, with a Noul question about reaching the goal riding in the same call. Top-k over summed log-probabilities makes it a beam search, drawn in an interactive graph.
ChoiceNoul - Project
typesafe-snake
Snake auto-played by Jev: one Choice per game tick, with legal moves and board facts computed in code before the call.
ChoiceReal-time control - Project
zadescoxp/Jev-Trades
A Next.js dashboard that streams one-minute crypto candles from Yahoo Finance and applies TypeSafe decisions to a simulated paper-trading portfolio, gated by a confidence threshold per risk profile. No live broker or order API is connected.
- Project
AI Elo Ranker
Recursive tournament engine that ranks texts such as poems, startup pitches and cold emails using Jev for pairwise judgments, Elo rating mechanics and Swiss matchmaking, with real-time WebSocket streaming.
- Project
tsai-sc (Jev plays StarCraft)
Harness that lets Jev control the original StarCraft shareware through synthetic keyboard and mouse input, recording action probabilities for each decision.
ChoiceReal-time control - Project
TypeSafe AI Playground (jev.works)
Community playground for Jev with around 110 use cases, games, dilemmas and model challenges. Supports editable prompts, A/B comparisons, conversation routing, document field extraction and code-policy inspection in a Next.js interface. Hosted at jev.works.
- Project
johnhughes3/LegalForecastBench
Benchmark that tests whether models can forecast federal motion-to-dismiss rulings from the judge's written record, scored with claim-defendant micro-Brier metrics and clustered intervals. Ships a Jev integration under integrations/jev alongside the frontier-model workflows.
- Project
GiesN/typesafe-jev-workflow
Async LangGraph workflow that sends an email to Jev, gets back a typed Choice of invoice or general, and routes it to a demo handler. Ships 10 labeled mock emails, and an API failure stops the run rather than assigning a fabricated intent.
- Project
frostney/clean-code-review
Judges every code file in a pull request against Robert C. Martin's Clean Code with Jev, one calibrated probability per question, then has a writing model turn those findings into two sentences per file and a decision for the change. Runs on Vercel's eve framework as one Next.js project.
- Project
raihankhan-rk/diffjury
Paste a public GitHub pull request URL and one Jev call returns risk bars and a verdict. The single request covers merge risk, review depth, whether design or security review is needed, missing tests, docs debt and how wide a bug could spread.
- Project
heist-one
Browser stealth game where Jev makes typed guard judgments while deterministic game code owns all world state.
- Project
anxkhn/JevPlaysPokemon
Jev plays Generation 3 Pokemon through Showdown and a real FireRed ROM, with no screenshots involved. Each turn the host reads the battle out of RAM and hands Jev the legal moves and switches; the author points a deliberately underlevelled party at the Elite Four.
- Project
kitze/pagegrade
Grades the sections of a web page for clarity, writing and on-page SEO with Jev, returning a score per section. Built as a browser extension with WXT.
- Project
morcoan/JevSeek
A local desktop and CLI coding agent where Jev selects the next tool from the user's request and recorded execution results, while DeepSeek or a local model supplies the arguments.
- Project
phureewat29/got-jev
Game of Thrones roleplay where a story model writes each scene and Jev reads it back, answering five questions in parallel: where Jon stands, the kind of beat, the danger, what should play under it, and whether the prose stayed inside the fiction. The header, soundtrack, artwork and next prompt are all functions of those answers.
- Project
zhuyansen/jev-search-rerank-eval
Graded relevance eval of a Jev score rerank against BM25, bge-m3 and other rankers over the Agent Skills Hub catalog, with 164 queries, 9,831 labelled pairs and the judge-circularity bias measured. Jev alone does not beat a good embedding ranker, but fusing the two wins even with Jev removed from the judging.
ScoreSemantic reranking - Project
jev-doom-agent
Browser-native Doom agent using a Chocolate Doom WebAssembly runtime, structured spatial state, composable AI controls and live decision telemetry.
ChoiceReal-time control - Project
lbotinelly/jev-little-airways
Toy archipelago where Jev makes every in-flight call: divert or continue, declare an emergency, give way, hold, and who lands first when two aircraft want the same runway. Each aircraft knows only itself and nearby traffic, asks four batched questions about every 1.6 s, and gets answers in roughly 150 ms.
- Project
TokenTrim/jev-agent-failure-benchmark
Benchmarks Jev on the 6,257 text traces of the Who and When Pro agent-failure-attribution set, predicting the responsible agent, the decisive step and the error type. Scored with the official whowhen_eval scorer against the paper's LLM baselines, at about $1.28 for the full run.
Choice - Project
Gaurav-Gosain/jev-sec-bench
Two blind security benchmarks for Jev on public corpora: 662 labelled prompt-injection messages and 200 matched vulnerable-code pairs, run through jev-go with raw per-sample output committed. At a plain 0.50 cut the injection run reached 96.5% accuracy, ROC-AUC 0.9927 and p50 325 ms.
- Project
jev-rerank-bench
Benchmark comparing Jev against Cohere Rerank 4, ZeroEntropy zerank-2 and a chat-model baseline across 14 datasets, publishing every raw API response and bootstrap ranges on each gap.
- Project
tsai-civ2 (Jev plays Civilization II)
Experimental full-game harness where Jev plays the original Civilization II in a browser, displaying live action probabilities.
ChoiceReal-time control - Project
ably-labs/jev-pong
Pong where the ball moves one segment per model decision, so a lane's speed is that model's latency and nothing else. Four lanes replay a recorded run with Jev, Gemini 3.8 Flash, Claude Haiku 4.5 and GPT-5.6 Sol on the right paddle, every player and agent on an Ably channel.
- Project
agent-labs-dev/fastbrowse
Browser agent that indexes the page into candidate actions and has Jev pick one, so it cannot click something that was never on the page, while an LLM plans and reads. Every claim in an answer cites a quote stored verbatim from the page, and Jev checks the claim against its quote.
- Project
grmkris/robo-harness
SO-101 robot-arm workbench with a Bun and Effect coordinator, a React workbench around Rerun, and a Python LeRobot hardware boundary. Its decision runner lets Jev, reached through the AI SDK evaluate API on Vercel AI Gateway, pick bounded joint steps under a spend cap, and chat and a decision run never move the arm at the same time.
- Project
milanboers/jev-plays-pokemon
Autonomous Pokemon Red agent on a PyBoy emulator. Jev reads a compact text snapshot of the game state, answers parallel yes/no questions each turn, and deterministic code turns the answers into button presses. Jev has no vision and no history, so the harness holds the short-term memory.
- Project
teyhouse/jev-secret-detection
Measures how well Jev spots real secret credentials in file snippets. Each case is one Noul question and the probability is compared with the expected label, with no regex matching or provider verification. Reports accuracy, AUC, recall, precision, Brier score, review-band counts and latency.
Noul - Project
vclic/smoking-extraction-benchmark
Paired comparison of Jev and OpenAI structured outputs on 1,000 synthetic outpatient notes, both given the same note, policy, candidate values and ten typed questions. OpenAI got 98.7% of records fully correct against Jev's 92.4%, while Jev cost about 133 times less with 5.2 times lower mean latency.
- Project
anisselbd/jev-phishing-bench
Reproducible comparison of Jev and Claude Haiku 4.5 on 2,000 phishing emails, asking whether an email agent should click the link. Haiku wins on accuracy, 81.3% against 62.6%, while Jev is faster and cheaper, and a logistic regression over Jev's five signal questions reaches 95.1%.
- Project
asfarsadewa/human-compiler
Paste an email, Slack message or memo and get rustc-style diagnostics back. A plain lexer finds phrases and their line and column positions, then one request to Jev scores dimensions such as passive aggression, unnecessary urgency and how much actual information the text carries.
- Project
bitnovus/jev-spam-eval
Asks how far a written decision rule gets you before you need a labeled training set. On 18,514 emails a written definition of spam scored 98.3% against 98.4% for TF-IDF logistic regression, and averaging both scores reached 99.2%. Also covers learning curves and distribution shift.
Noul - Project
cephalization/jev-triage
Multiplayer triage dashboard for public GitHub repositories, where a System One model answers a fixed set of typed questions per issue: what kind, how severe, how urgent, whether it duplicates another, and what a maintainer should do next. People correct the answers, corrections are shown back on later runs, and nothing is written back to GitHub.
- Project
citation-verifier
Checks whether each cited paper actually supports the sentence citing it: Claude locates the quote, Jev scores the support, a human makes the final call. Live at verify.papertrellis.com.
- Project
colinmcdermott/emoji-jev
Emoji autocomplete that makes one Jev call per pause in typing: a Choice over up to 254 emojis, a Choice over eight emotions, three Scores and three booleans, answered in parallel in about 100 ms of model time per the README. TanStack Start on Cloudflare Workers through the AI SDK typesafe-ai provider, hosted on Whop, with a live demo at emoji-jev.whop.site.
- Project
Friedjof/jev-mobile
Proof of concept that turns a high-level Android goal into a bounded observe, stabilize, decide, act loop. Jev receives a compact semantic UI state and a short list of already-valid actions, and never generates coordinates, MCP calls or code. Exercised against Android Settings on a real device.
- Project
gtaras7/typesafe-jev
Local CV screening workbench that asks Jev a small fixed set of typed questions about each CV in a folder, scores the answers with arithmetic you can read, and shows a sortable shortlist. Judgments are stored apart from the arithmetic, so editing the role policy re-scores every candidate in about 20 ms at no cost.
- Project
hegargarcia/jev-playground
Benchmarks Jev against Luna, Haiku and Gemini at choosing the next action in games with explicit states, legal actions and measurable outcomes. Each game is a state machine where code owns the rules and the model only picks among the options it is given, with every model reached through Vercel AI Gateway.
- Project
jev-gomoku
MoonBit client for the Jev API plus a Jev-vs-Jev gomoku match with per-move timing logs. Write-up at zenn.dev/mizchi/articles/jev-plays-gomoku.
ChoiceReal-time control - Project
jev-play-ping-pong
Jev plays a browser table-tennis game in real time from structured telemetry, driving ordinary Chrome inputs and logging auditable decision evidence.
- Project
jev.mods
Minecraft mod in which Jev attempts to finish the game from scratch with no scripted route.
- Project
jgridifier/jev-research-eval
Reproducible eval harness and field note for a Jev Ultrafast research-browser session: 11 baseline cases, human and quant stress suites, QC grades, a suite runner and a report generator. It drives an upstream checkout of browser-use/jev-ultrafast at a pinned commit rather than forking it.
- Project
joshlarsen/jev-t-rex-runner
The Chromium T-Rex Runner game with Jev at the controls: it picks jump, duck or keep running for each new obstacle and chooses a short or full jump profile. The browser keeps speed-aware frame timing, collision geometry and input execution, and manual play still works without a key.
- Project
jourdanlabs/assay-001
Pre-registered independent check of Jev's calibration and type-safety claims, with the protocol frozen and hashed before the first call. On CLINC150 the chosen-option probabilities were calibrated (ECE 0.0204), on Banking77 they were not (ECE 0.0936, systematically overconfident), and across 8,576 responses there were zero type errors.
Choice - Project
kxzk/typesafe-jev-drone-demo
Three.js drone simulation with a typed Python and FastAPI backend where jev-latest chooses each navigation action. The Choice distribution, Noul obstruction estimate, confidence, latency and the exact model input and output are visible beside the flight.
- Project
santos-sanz/jev-audio-beeper
Proof of concept that identifies Spanish profanity with Jev and replaces the matching word intervals with a beep without changing the audio duration. Word timestamps from an ASR adapter feed parallel Jev evaluations for a Boolean profanity probability and a severity Score, then ffmpeg overlays the beep and writes a JSON audit report.
ScoreNoul - Project
siroccomask/snake-jev
Desktop Snake where Jev assesses the board and Python combines the answers into a move, one API call per game tick. One recorded run reached 29 foods in 461 ticks before trapping itself; the agent spots immediate hazards but does not plan a route.
- Project
TarunTomar122/jev-askable-arm
Zero-shot robot tasks on a simulated Franka Panda in ManiSkill. Given a plain English goal and the scene, Jev picks one of about 30 hardcoded primitives per step and Python runs it, at roughly one second per pick. No training, demos or policy.
- Project
jev-headline-bench
Tests whether Jev can pick the winner of real headline A/B tests across 10,984 Upworthy randomized experiments. Reported 64.5% overall and 74.7% on decisive pairs.
- Project
typesafe-minecraft-demo
Minecraft Java player controlled by Jev, with live decision display, a Canadian flag building task, and a side-by-side dashboard.
ChoiceReal-time control - Project
SIEGE (CoreWeave Hacks: Agent Loops 2026)
Hackathon project: 200 people versus one agent behind a typed System One action gate that learns from each breach, evaluated with W&B Weave and hardened by a defender loop. Playable at vnmoorthy.github.io/siege.
- Project
Browser Use Olympics + almond-fastloop
Five-task timed benchmark for browser agents plus almond-fastloop, a roughly 200-line agent wiring Chrome DevTools state to a bounded Jev decision and an executor. Leaderboard as of 2026-09-17 shows fastloop at 13.8s total versus Codex 66.2s and Claude (Cowork) 100.3s.
- Project
Jev Driving Lab / typesafe-playground
Three runnable experiments: support-message triage producing six independent judgments, a driving simulation where Jev picks lane and target speed against traffic and signs, and an alternative Fable implementation. Shows questions and probability distributions for each call.
- Project
jev-bfs (Wikipedia link races)
Wikiracing solver that uses Jev to rank outgoing links by fewest expected remaining hops, scoring up to 128 candidate links per request with up to five concurrent requests, shown live in the terminal.
- Project
jev-rerank
Uses Jev as a calibrated reranker: one relevance question per document, up to 30 documents per call, returning a probability usable both as a ranking score and an absolute threshold. Apache-2.0; benchmarked on SciFact, NFCorpus and FiQA. The repo now redirects to hev/reranker.
- Project
Jev for AI Agents: I tested it with rtrvr.ai for browser automation
Uses Jev as the decision layer for browser automation with rtrvr.ai and shows the results.
- Project
TypeSafe AI Jev (SystemOne model) plays Kyrandia 2 against itself: running in circles
Community experiment driving the adventure game Kyrandia 2 with Jev choices; the model ends up looping. A useful failure case for real-time control.
AI Adventure XP40OpenTypeSafe AI Jev (SystemOne model) plays Kyrandia 2 against itself: running in circles on youtube.comChoiceReal-time control - Project
Jev Tac Toe
A browser tic-tac-toe game where Jev picks its move each turn with the choice primitive over the open squares. A move journal shows the probability Jev gave each square, a blitz mode gives the human one second per move, and every finished round joins a shared scoreboard.
Josh Pigford (@Shpigford)OpenJev Tac Toe on jevtactoe.neato.funChoiceReal-time control - Project
Jevsweeper
A live Minesweeper broadcast where Jev picks every reveal and flag with the choice primitive and everyone watches the same game. A decision log shows each move's confidence and alternatives, a separate checker labels each loss as a forced guess or a missed deduction, and the game pauses when nobody is watching.
Josh Pigford (@Shpigford)OpenJevsweeper on jevsweeper.neato.funChoiceReal-time control - Project
I built a side project to test TypeSafe's Jev model with no generative LLM that scores your project idea
A side project that scores a project idea using only Jev Score questions, with no generative model in the loop.
ScoreComposite scoring - Project
Companion, a Jev demo
A 3D robot companion driven by Jev: one call per turn answers nine typed questions, and code decides whether to act, ask or shrug. The panel shows the tracked mood and what the robot decided.
jev-demo.vercel.appOpenCompanion, a Jev demo on jev-demo.vercel.app - Project
Crowdcheck
Live demo that tests a post against a crowd of 10,000 AI readers with their own jobs, tastes and memories. Jev answers whether each one likes, dislikes, reposts, follows or blocks, and every reaction comes from an AI reader, not a real person.
crowdcheck-ai.vercel.appOpenCrowdcheck on crowdcheck-ai.vercel.app - Project
Hollow Creek
Browser village where nobody follows a script. Each villager watches what you do, remembers it and decides how to treat you, answering with expressions and actions rather than talk, and a side panel shows the state sent to Jev, the questions asked and the answers returned.
hollow-creek-sigma.vercel.appOpenHollow Creek on hollow-creek-sigma.vercel.app - Project
Jev plays Pac-Man
Pac-Man recreation steered by Jev from structured game state. Before each junction the game sends the maze, the actors and the legal directions, and Jev answers with one direction and a probability for each option, with no screenshots, no scripted bot and no hidden pathfinding.
jev-pacman.ephraimduncan.comOpenJev plays Pac-Man on jev-pacman.ephraimduncan.com - Project
Moderation Guard
Moderation playground that screens text in one System One request: parallel Noul probabilities for profanity, hate, sexual content, violence, illegal help, self-harm and spam or jailbreak, plus a severity Score. The verdict comes from policy thresholds in your code, review at 0.35 and action at 0.70.
guard-jev.vercel.appOpenModeration Guard on guard-jev.vercel.app - Project
Room, powered by Jev
Interactive 3D room that reacts to a few words: Jev sets lighting, curtains, music, TV, volume and the desk lamp, and the app renders the result. Shows the raw API response and an estimated per-call cost, with presets that make no call at all.
jev-room.moe136231.chatgpt.siteOpenRoom, powered by Jev on jev-room.moe136231.chatgpt.site - Project
Tetris x TypeSafe
Tetris you can play yourself or hand over to the TypeSafe Choice primitive, which picks every move. Score, lines and level are tracked while it plays.
jev-omega.vercel.appOpenTetris x TypeSafe on jev-omega.vercel.appChoiceReal-time control - Project
Tweet Tester: Preview & Score a Post Before You Send It
Free SuperX tool that scores a draft X post from 0 to 100 against a normal post from an account your size, with lists of what helped and what hurt and a simulated feed preview. Rob Hallam's launch post says each draft gets 61 Jev questions in about a second, fitted on 9,481 posts from 207 creators.
Rob Hallam (@robj3d3)OpenTweet Tester: Preview & Score a Post Before You Send It on superx.so - Project
TypeSafe Typewriter
Live Val Town demo: type a sentence and watch 16 calibrated judgments move on every keystroke, including passive-aggressive, AI-written, tone and urgency, all from one API call.
typesafe-demo.val.runOpenTypeSafe Typewriter on typesafe-demo.val.run - Project
Yes / No
Free demo of Jev with no sign-up: ask any question and get Yes, No or Maybe, grounded in live web search when it matters. Calls jev-latest through TypeSafe's System One API.
Nothing matches those filters
Clear a filter to widen the search, or send us the thing you were looking for.