Awesome Jev

What people built with Jev (TypeSafe System One). Showcase, not a tools directory. Each line is a project and what Jev does inside it.

Routing

• Foreman — Jev on a Codex worker for semantic supervision on a factory floor
• llama-index-jev — LlamaIndex reranker + router with typed Jev scores/choices.
• opencode-jev-orchestrator — Sticky cheap model; Jev escalates hard turns to stronger subs.
• agent-router — Picks Cursor/Claude/Codex/OpenCode + effort, then launches.
• todo-jev — Task classifier + 3-tier routing engine powered by Jev.
• tiershift — Every LLM call → cheapest capable model; Jev routes ~ms.
• jev-router — Cost-aware Choice: cheapest model that can handle the query.
• JevRouter — Routes models / tools / subagents with Jev Choice.
• hono-jev-router — Hono semantic router: HTTP requests routed by meaning via Jev.
• pi-jev (Theo) — Pi: semantic tool routing + typed System One decisions.
• flue-jev-demo — Flue agent routing with Jev through Cloudflare AI Gateway.
• Jev-Auto-Router — Codex per-call GPT model routing via Jev Choice.
• jcm-router — Local proxy: Jev picks Claude model + effort per message.
• pi-jev-router — Pi: Jev Choice routes model via Vercel AI Gateway.
• jev-router (prismhq) — LiteLLM proxy: Jev Choice picks which model runs next.
• tax-doc-classifier — Choice IRS form + page class; code acts only when confidence ≥0.95
• Codriver — per-turn Choice picks which fleet model fits this turn, then rewrites the outbound message
• ReflexRoute — zero-shot / few-shot Choice routes which model runs next
• jev-router — per-turn routes Claude Code / Codex to cheap vs strong model while keeping each CLI’s native session
• SemDecide — Jev as a Unix CLI: classify, score, filter, and guard in scripts and CI
• Jev Codex Router — Jev scores how hard each coding turn is, then picks a cheap vs stronger model
• safer-with-jev — Jev classifies the task, then routes it to Grok or GPT (was typesafe-on-neon)
• Lahfir harness router — Jev routes Claude Code / Codex / Opencode tasks by intensity (Haiku vs Opus sub-agents vs external harness)

Refund review

• (none in this batch)

Realtime

• jev-ultrafast — each step Jev picks the operation and the element from DOM state; small LLM only types; flight search demo 7s / $0.0039; demo
• tsai-civ2 — Civilization II in browser: live Jev action probabilities.
• jev-triage (issues) — Issue triage across large repos with typed Jev judgments.
• jev-chat-for-twitch — Chrome: filter live Twitch chat with Jev.
• jev-block-android-ad — Android: JevNoiseGate filters notifications/SMS by meaning.
• jev-robotics-demo — MuJoCo robot arm: Jev vs Claude Opus decisions.
• jev-reflex-autonomy-lab — Multi-drone autonomy: Jev reflex decisions.
• typesafe-jev-drone-demo — Three.js drone: live Jev navigation choices.
• jev-paint — Use Jev to make art (closed action choices).
• jev-pong — Pong: ball moves one step per Jev/LLM decision.
• snake-jev — Snake: parallel Jev assessments, one API call per tick.
• tsai-sc — StarCraft shareware: Jev controls keyboard/mouse.
• jev-plays-pokemon-red — PyBoy Pokemon: code owns route; Jev picks at branches ~100ms.
• jev-plays-pokemon — Pokemon Red: Jev picks actions.
• jev-bfs — Wikipedia link races: Jev ranks next hop live.
• jev-canvas — Voice + finger on tldraw; Jev decides draw intent.
• jev-desktop — Codex Computer Use: Jev selects the next UI action.
• jev-for-chrome — Chrome: Jev drives the tab you’re looking at.
• jev-agent-browser — Bounded browser agents: typed actions via Jev.
• jev-browser — Browser use with Jev as the decision model.
• fastbrowse — Jev picks each on-page action; LLM reads/plans around it.
• jev-skip — Caption-time sponsor skip: probability per span → skip.
• jev-triage — Choice case/desk + Score priority + Noul fraud/recall; code sets queue/SLA/next step
• Jev-MongoDB — Choice category + Score mood/urgency/safety on Change Streams; code writes back and pushes a board
• youtube-sponsor-detection — Choice over caption lines / Noul “spoken ad?”; code maps timestamps and skips
• jev-voice-browser — spoken partial transcripts → Jev picks intent/target (~250–350ms); Playwright acts; text spans copied verbatim
• WebMCP × Jev bench — Jev picks WebMCP tools; Mercury fills args; 49/49 at ~112×–245× lower model cost vs GPT-6 Astra computer use; thread · WindTunnel
• 1v1 Jev — Jev at about 9Hz picks move, aim, ADS, shoot, and jump
• Playground — Jev experiments: PHI detect, comment review, live tone, job and industry classify
• Prism — Jev scores toxic flow, market stress, mean-reversion, and liquidity (shadow only, no live trades)
• neo4jev — Jev scores candidate graph edges, then beam-searches the likely path
• jev-desktop — Jev reads the Accessibility Tree and picks which control and which action
• TypeSafe Mario — Jev chooses NES controller macros from structured emulator state (Choice + Noul + Score); no screenshots
• TypeSafe Computer Use — OCR reads the Mac screen; Jev picks the next UI action (~$0.0002/step); demo thread
• Kyle Jeong + Stagehand — a11y tree as state; Jev picks the next browser action, Stagehand executes; ~$0.001/task, near-instant remote browser
• jev-trader — Jev watches the Kuru MON-USDC book and posts a post-only limit quote each Monad block (~300ms)
• Sprite Fusion realtime levels — game state in; Jev Choices for next terrain slice (type/width/gap/height); code places blocks (~319–375ms, ~$0.00057/request)
• jevpilot — Choice among sampled drive paths at ~1.5–4Hz in a Three.js driving sim
• jev-drone — classical CV builds the scene; Jev picks maneuver / risk / lost-target at ~2.5Hz in MuJoCo
• mobile-jev — Jev chooses each Android UI action; Uber demo ~21s / 9 actions (to checkout, no purchase)

Eval

• Jev Review — Jev pre-screens correctness, security, reliability, compatibility, and test risk before a heavier model
• jev-architect — Find, design, and evaluate TypeSafe Jev decision loops.
• jev-classifier — Local tool-routing classifier for coding agents + gateway.
• pi-typesafe — Pi: batched evaluation tool + typed API for extensions.
• jevvy — Jev-powered plugins for coding agents.
• jevwire — MCP + DecisionModel library + escalate-only Claude hooks.
• jev-use — Plugin: hand no-text-output steps to Jev judgments.
• jev-skill-suggester — Bounded installed-skill recommendations via Jev.
• jev-spam-eval — Zero-shot spam filtering with Noul vs TF-IDF baselines.
• mastra-jev-moderation — Input moderation for Mastra agents via Jev.
• jev-auto-approve — Jev PR auto-approver.
• jev-sentinel — Pi: Jev checks tool calls/outputs/replies (injection etc.).
• jev-security-scan — Review Skills/MCP code for suspicious behavior with Jev.
• jev-review (Niaz) — Local-first continuous quality review MCP for coding agents.
• perch — Semantic code linting against configurable Jev questions.
• abide — Make coding agents abide by project rules (Jev-backed).
• pi-quiet-ask — Pi quiet decision layer via TypeSafe Jev.
• stanley-code — Bounded TypeSafe Jev workflows for coding agents.
• limpet — Stop hook: plain-language rules so agents don’t stop too early.
• wakegate — Ask Jev if a sleeping agent wakeup is worth a full LLM turn.
• jev-belay — Claude Code Stop hook: evidence check before trusting “done”.
• diffjury — PR risk router + code-review coach via Jev.
• jev-logtriage — Choice whether a log batch is worth acting on; confidence gates.
• is-malicious — Scans source/CI files; Jev flags suspicious behavior.
• hunch — Semantic code review with plain-English rules + Jev.
• jev-pref — Lints code changes against AGENTS.md prefs via Jev.
• hermes-jev-approvals — Hermes smart command approvals reviewed by Jev.
• jev-commit — Pre-commit: one Jev call — message match staged diff?
• jev-git — Pre-commit / pre-push semantic reflex gate.
• pi-verdict — Minimal Pi permission gate (auto-mode style).
• jev-guard — Risk-scores every tool call → allow / ask / deny.
• pi-jev — Pi tool-call gate + jev_ask for typed calibrated answers.
• pi-jev-permit — Pi bash/write/edit calls: Noul permit gate before run
• jevmod — multi-category Noul moderation (spam/scam/…) with thresholds you own
• Paolo Rosson PR review — paste a diff → one Jev/TypeSafe call → ~14 typed checks as probabilities → BLOCK / security review / nits / merge; ~$0.00007/PR
• fx auto-review — Jev as the safety classifier for fx auto mode; ~5–18× faster and more accurate than GPT-5.6 Luna on their benchmark
• pi-jev-auto-mode — semantic allow/block on Pi bash/write/edit; fails closed when undecidable (~193–642ms)
• pi-warden — scores write/edit vs project rules plus irreversible/off-task on tool calls; steers the Pi agent
• Jev-Moderation-Bot — Choice LEGITIMATE|SPAM|SCAM_LINK plus Noul ban-now? per Discord message
• triagedy — per alert: disposition close|escalate|contain|investigate plus severity/FP/IR probabilities; Rust policy maps to action
• privacy-facts — ~14 privacy-policy questions as typed labels + confidence; picks supporting policy sections for excerpts
• sokit — control questions each turn (next_action / spam-route / draft-safety Noul); ConfidenceGate → act|confirm|escalate

Writing check

• Human Compiler — Jev scores prose for information density, clarity, hostility, and corporate tells, then the compiler emits rustc-style diagnostics
• Jev-X-Sentiment-Analysis — X posts: Jev sentiment / judgment labels.
• slop-grader — Rule-based slop grader for text files powered by Jev.
• pagegrade — Choice/Score page sections for clarity, writing, SEO.
• JevSlop — Jev judges whether a note/article is AI slop.
• taste-lint — Catch AI slop before you ship.
• snifftest — Prose linter for AI writing tells (rules + optional Jev).
• draftpulse — Choice draft category + multi Score (hook/specificity/reply/share/dwell/slop); code weights into a label
• script-judge — Choice approve / revise / reject on a pasted script (does not rewrite it)
• Jevibe Check — Jev labels drafts and posts for warmth, constructiveness, tension, sarcasm, clarity, and intent

Tools

Official SDK / skill / Gateway / docs: jev.gallery/tools

• typesafe-mcp — Jev as Choice / Score / Noul inside Claude Code, Desktop, and Codex
• jev_jsonschema — Run a JSON Schema through Jev API; get JSON judgments back.
• jevkit — Rust CLI for typed decisions + offline lint before you pay.
• jev-cli — Small dependency-free CLI for TypeSafe Jev judgments.
• decide-mcp — Configurable decision MCP with Jev + bias profile routing.
• jevalyn — Rails-native wrapper around Jev System One API.
• jev-dsh-decision — Structured decision plugin for agent harnesses.
• ask-jev-skill — Hermes/agent skill to ask TypeSafe Jev.
• jev-workbench — Versioned judgment functions on Jev; call published versions.
• jevcal — Calibrate/threshold/drift-check typed decision models.
• jevocks — Everyday stocks status judgments with Jev.
• jev-trade — Live Hyperliquid trader driven by Jev choices.
• jev-cvss — Fast CVSS scoring from vuln descriptions via Jev.
• jev-web-analyzer — See what Jev thinks about a SaaS website.
• jev-scout (MCP) — MCP research: every query/result/page scored by Jev.
• advocaat — Type-safe client for asking Jev questions about datasets.
• jev-mcp (blakestone) — MCP: classify/score/check/match/screen tools.
• jev-mcp — MCP: put TypeSafe Jev on Cursor/Codex coding loop.
• Jevbridge — ACP/MCP adapter: Jev beside coding/chat models.
• grok-bot-jev — Grok Bot decision layer: usage gates + Jev skill templates.
• hunch (Ruby) — Ruby/Rails probabilistic control flow via Jev likely?.
• sqlite-jev — Batched NL judgments for SQLite via Jev.
• jevql — Semantic SQL for Postgres powered by Jev.
• jegrep — Semantic grep: find code by describing what you want.
• jgrep (keltokhy) — Filter lines by meaning with Jev (description as pattern).
• jev-curate — High-throughput dataset sifter powered by Jev.
• jev-cookbook — Practical Jev recipes (triage etc.) on OpenRouter.
• new-api-plugin-typesafe — new-api plugin: native /v1/systemone Jev tasks.
• jev-agent-skill — Offload classify/screen/score/verify to Jev from agents.
• jev-judgment — Agent skill: closed coding judgments → TypeSafe Jev.
• super-jev — Small decision-to-action harness for TypeSafe Jev.
• pi-typesafe-jev — Pi extension: five tools wrapping TypeSafe Jev judgments.
• jev-engineering — Decision layer for agents: hard rules then Jev call.
• jev-mcp (BYK) — MCP: typed noul/choice/score judgments for agents.
• jev-scout — Repo/crate scout: Jev relevance without hallucinated claims.
• jev-seo — SEO/GEO CLI + MCP: Jev-backed ranking/judgments.
• jev-social — IG/TikTok/LinkedIn research: typed routing + evidence.
• fast-dev-compaction — Codex plugin: Jev-guided context restoration around compaction.
• pi-fast-jev-compaction — Pi: verbatim compaction keep/truncate/drop via Jev.
• yoshi — Context-pruning proxy: Jev judges which history is still needed.
• jev-pruner — Trim long Bash output with Jev before the model sees it.
• jev.nvim — Neovim: Treesitter splits; Jev scores functions → quickfix.
• jsort — Sort lines by meaning via pairwise Jev comparisons.
• jgrep — Semantic code search: what code does, not what it’s called.
• jev-nlgrep — Natural-language grep powered by Jev.
• jev-semgrep — Semantic grep: Jev scores each line against a meaning.
• jev-reranker (shinpr) — Rerank/filter/compress JSON search hits with Jev.
• jev-reranker — RAG relevance filter/rerank with Jev probabilities.
• citation-verifier — Noul: does each cited paper support the citing sentence?
• jev-axi — CLI: pick / rate / check / rank / triage / guard.
• docjev — Doc Q&A / retrieval judgments with Jev.
• sift — Chrome: label every X post (substance/humor/promo/junk/…).
• typesafe-jev — Screen a folder of CVs with typed Jev judgments + policy.
• jev-rules — one Noul per standing rule (“is this request about that?”); Claude Code only sees matching rules
• pi-jev-skill-suggestion — Pi extension: strip skill listing; Jev gate + wide/narrow Choice → at most one skill (cookbook two-stage; fail open)
• classifier.dev (repo) — hosted zero-shot labels via Jev (fast=Jev; smart=unsure → reasoning model)
• skillbox — self-hosted skills library (MCP); optional Jev recommends which skill fits the task
• jev-shell-history — zsh autosuggestions: Jev ranks which past command you are completing as you type
• pg-jev — Postgres extension: filter/rank/classify rows with plain-language Jev conditions in SQL
• unclutter — browser extension: Jev identifies page clutter/ads and hides them via reusable template rules
• notra — Choice sentiment plus rank/position classifiers for GEO / AI-visibility mentions
• HA-Jev — Noul/Choice/Score on Home Assistant house state → sensors; Assist routes by confidence
• jev-mcp — Jev as MCP tools: verify claims, screen prompt injection, rank by meaning
• Blink — at each directory level Jev scores which files or folders matter, then sends more walkers there
• Winnow — Jev judges which Read / Bash / Grep output is actually relevant before it enters Claude Code context
• fast-jev-compaction — Jev scores each tool call/result (keep / truncate / drop) for Claude Code compaction; demo
• SkillRanker — Jev ranks which agent skills fit the next step from live session context (Choice + Noul; abstention when none fit)
• jevmeter — scores every transcript sentence for BS/spin/hype; ffmpeg burns a live meter into a 16:9 edit
• typesafe-adblock — code finds ad-shaped DOM nodes; Jev Noul per candidate “is this a paid ad?”; removes above threshold (BYOK Chrome ext)
• jev-search — Jev picks sources/time/query, then scores hit relevance; returns ranked links/snippets only; live
• typesafe-skill-router — Jev names the one Hermes skill worth loading before the model call (~$0.001/turn); abstains if low fit
• jev-agent-skill-router — batched Choice over a skill catalogue → route / no_skill / review with need/ambiguity/fit Nouls
• typesafe-mod — Claude Code hook — Jev ranks installed skills per prompt; can also score AskUserQuestion options in parallel
• OpenRouter — Jev as a System One decision model on OpenRouter (beta); typed state + question → typed decision with probability; announcement
• LangChain + Jev harness — TypeSafeClassifier / middleware: model routing and Auto Mode safety checks before tool calls; post
• jev-experiments — latency-focused Jev demo suite (agent-assist, commit-sentry, shell-guard, dispatch, …): multi-question typed judgments in ~100ms
• Six things with Jev — Isaac Flath: fact-check scripts, rank news, find PDF text, check citations, group review notes, eval agent-failure traces; thread
• cursor-clijev-compaction — same keepCall/keepResult nouls for Cursor CLI agent history; re-injects kept facts after native compact
• Brainwires jev-mcp — MCP tools jev_rank / jev_verify / jev_next_step / jev_gate_action for chunk ranking, claim check, turn-done, and pre-destructive gates
• Reflex — in-browser WebGPU Qwen doing noul/choice/score with probabilities; demo · thread
• json-render + Jev — Jev chooses catalog components / slots / action bindings; code assembles the UI spec (no free-form JSON); docs · demo
• WTF is Jev (9 builds) — Matt Van Horn roundup of early builds (browser agent, compaction, fx, routing, RAG, games, email, batch eval, on-device); thread
• RAG chunk filter — per retrieved chunk, keep vs drop for precision (no author repo yet)
• Minecraft realtime — Jev for fast realtime actions while a planner LLM plans (multi-zombie fight)
• Email triage batch — batch inbox triage/classify/route (Vogel demo; ~1500 emails in batches of 100)
• Every mini vibe check — 21 AI-writing checks × 37 docs → 777 judgments in under 0.7s (~$0.0025)