LLMs write essays. Jev makes the call. This is the community's living index of what people are actually shipping with TypeSafe AI's System One decision model: real repos, real measurements, and the patterns that already work.
Jev is TypeSafe AI's first System One model: a decision engine that returns typed, calibrated judgments (Choice / Score / Noul) instead of free-form prose. It runs in 70–500 ms at ~$0.042 per million input tokens, and output tokens are free. Fast enough to sit inside a game loop, cheap enough to run over millions of rows, and constrained enough that your code can branch on the answer without parsing JSON.
Every listing links to a public source. Headline results are tagged [self-reported] or [independent], and every link is scored against a written rubric before it lands here (see CONTRIBUTING.md).
- What is Jev
- Official Resources
- Top Use Cases
- Community Builders
- 01 Workflow · 02 Bulk · 03 Realtime · 04 Verify · 05 Harness · 06 Voice
- Benchmarks
- Popular Blogs
- Contributing
- License
You send Jev a state (a string, JSON object, or list of text) plus typed questions. It returns typed answers with probabilities: no invented fields, no malformed JSON, no off-schema values. Your code owns the control flow, the thresholds, and the side effects.
| Primitive | You declare | Jev returns |
|---|---|---|
| Choice | A list of options (up to 255) | The chosen option + probability per option + confidence |
| Score | An ordered rubric (2–10 levels) | A probability-weighted score + confidence |
| Noul | A yes/no statement | A single probability from 0 to 1 |
Three primitives is the whole API. That's the point. Questions in one request are evaluated in parallel and in isolation against the same state, so asking ten questions costs roughly the time of one. That one property unlocks speculative fan-out, confidence-gated routing, and every pattern on this list.
The economics (vendor-reported): $0.042 / million input tokens, output free, 70–500 ms end-to-end. On TypeSafe's own four-workflow eval, Jev lands at 67.8% agreement with a frontier-model consensus at $0.0004 per case: essentially tied with GPT-5.6 Terra (67.9%) at ~1/76th the cost. Treat the multipliers as directional, not gospel: the honest independent numbers are lower.
TypeSafe (official):
- TypeSafe AI · Introducing System One Models & Jev (launch post)
- Docs · API reference · Primitives · Patterns · Models & pricing · Demos
- Playground · Workflow evals · Console
Official SDKs & tools:
- typesafe-sdk-python · typesafe-sdk-js · system-one-adapter-python (run LLMs over the same interface) · agent skill
Bypass the waitlist: the Vercel AI Gateway route (typesafe-ai/jev).
Community clients: Elixir · Ruby · Rust · .NET · Go · Scala/ZIO · PHP/Laravel
The use cases with the most traction since launch: ranked by popularity, not raw star counts (stars are a hidden ordering signal only; rigor does not track reach).
| Project | What it does | Headline result |
|---|---|---|
| browser-use/jev-ultrafast | Browser agent with a dynamic indexed action space; a small LLM writes text only when a step needs it | Zürich → London Google Flights in 7.1 s / $0.0039 |
| tamaratran/fast-jev-compaction | Claude Code compaction → Jev keep/delete | Content stays verbatim |
| jarrodwatts/jev-trader | Market maker on Monad/Kuru: one buy/sell decision per block | 81 ms model latency per block |
| awlevin/typesafe-computer-use | macOS computer use via OCR + Jev action choice | $0.0002 / decision vs Opus 5 at $0.032 |
| thruwire/foreman | Supervisor over Codex workers | Architecture experiment for software factories |
| devagrawal09/jev-review | Staged code reviewer + local dashboard | Risk matrix → evidence → severity → routing |
| fhshaik/typesafe-mario | Super Mario from emulator RAM as object-centric JSON | No screenshots sent to the model |
The "few upcoming": high rubric score but below the popularity cutoff. This tier is where rigor beats reach: several of these have single-digit stars and the strongest measured results.
| Project | What it does | Headline result |
|---|---|---|
| realZachi/pg-jev | Natural-language WHERE clauses for PostgreSQL: no embeddings, no vector column |
129 rows in ~1 s for ~$0.0009 |
| droidrun/mobile-jev | Android agent: Jev decides each tap | Uber route in ~21 s / 9 actions |
| DevMortimer/pi-warden | Agent guardrails that steer instead of interrupt | 6 rule breaks → 0 across 150 paired runs |
| AnshChoudhary/typesafe-ai-firewall | Pre-execution firewall for agent tool calls: one Noul per hazard | 0% hard negatives blocked vs 39.2% with a single "is this dangerous?" prompt |
Every headline result above is self-reported by the project author. Treat them as launch-week proof-of-concept signal, not production case studies.
The long tail of unique use cases, grouped by decision shape. Each entry answers two questions: what does it do, and why is it unique.
Smart if-statements inside ordinary software: the fuzzy middle between brittle rules and expensive chat models.
- kylemclaren/jevql:
WHERE jev(...)filters,jev_probsorts, andjev_choicegroups on a vanilla Postgres with no extension, judged client-side. Unique: SQL filtering without a plugin or embeddings. - AboveColin/HA-Jev: Home Assistant integration; typed questions over entity state become sensors and automation actions. Unique: a decision layer for household automations.
- valentynkit/jev-commit: one Noul on whether a commit message matches the staged diff, plus debug-leftover and unmentioned-work checks. Unique: pre-commit semantic lint instead of a full review.
- valentynkit/jev.nvim: plain-language buffer search; Treesitter splits the file into functions, Jev scores each, results land in quickfix. Unique: semantic editor search with no embeddings.
- DanRWilloughby/snifftest: prose linter that sniffs out AI-writing tells; countable regex rules run locally, judgment rules send one paragraph at a time to Jev. Unique: one probability per rule, no general model, and it never rewrites your text.
Cheap judgment over giant corpora: the economics that make "run it on everything" rational.
- valentynkit/jev-skip: sponsor-segment probability on YouTube's seek bar, scored from the caption track alone. Unique: crowd-free sponsor detection (77% of SponsorBlock's sponsor seconds, ~$0.0008/video).
- youkiti/tiab-review-plugin: title/abstract screening for systematic reviews across labeled medical datasets. Unique: evidence-synthesis triage at scale (95% recall @ 16,645 records).
- Canonical examples: 1,018 papers → 24 topics for ~$0.08; 2,000 wine notes → CatBoost numeric features at 1.77 RMSE.
Action selection at game, UI, and market clock rates: perception and safety stay in code, Jev picks the move.
- phyous/tsai-sc: StarCraft shareware harness with 421 structured decisions. Unique: real-time strategy from structured game state.
- valentynkit/jev-plays-pokemon-red: code owns the route, Jev picks only at branches, and predictions are scored by Brier against RAM. Unique: calibrated in-game decision logging.
- RomanSlack/jev-drone: camera-only quadrotor; Jev is advisory at ~2.5 Hz while safety stays in code at 50 Hz. Unique: split-rate control (tactical vs safety loop).
- Canonical examples: Doom (~10 Hz), Wikiracing, Snake, Tetris, Pac-Man, Subway Surfers ×50.
Score, judge, and gate prompts, traces, tool calls, and claims: at a fraction of the LLM call you're protecting.
- valentynkit/jev-belay: blocks an unverified "done" in a coding agent; spends one Jev call only when files changed with no passing check since, and fails open on every error path. Unique: evidence-gated task-completion guardrail.
- kiarina/labs: safety judgment: moderation plus shell-command safety checks. Unique: strong non-English (Japanese) moderation: 36 misses vs OpenAI's 292 on 826 harmful texts.
- Canonical examples: citation grounding, jailbreak and prompt-injection screening, pre-execution shell checks.
Make the agent loop cheaper, safer, and composable: Jev as the load balancer above models, tools, and humans.
- jkudish/jev-mcp: MCP server exposing claim verification, content screening, and semantic ranking to coding agents. Unique: verification + injection screening over MCP.
- itsmostafa/typesafe-mcp: single-binary Go MCP for Claude and Codex. Unique: zero-dependency agent integration.
- AbdelStark/bicameral: Pi harness: the LLM writes, Jev supplies typed reflexes for policy, loop detection, and review. Unique: a decision layer as the coding-harness reflex arc.
- Canonical examples: skill routing, context reduction, model routing.
Sub-second decisions on the audio path: the youngest pillar, still mostly conceptual.
- Spoken "go back" mapped to a click in ~300 ms (voice → action)
- End-of-utterance and turn-taking detection via a Noul over the voice-activity signal
- Speak-up gates for wake-word-free assistants
No canonical repo has shipped yet; this is the pillar to watch.
Rigor over reach. Everything here is tagged independent or self-reported.
-
Every: "Mini-Vibe Check" (Mike Taylor): 777 judgments in <0.7 s (~$0.0032 each), 6-of-7 planted defects caught, ~25× faster than Claude Fable 5.1.
[independent] -
Archer Hume: "Architecture Unmasked": 84.6% MMLU-Pro, ECE 0.0313. Third-party probe, not a TypeSafe figure.
[independent]
- Gaurav-Gosain/jev-sec-bench: blind prompt-injection & vulnerable-code detection: 96.5% accuracy, ECE 0.0588, 662 samples
- TokenTrim/jev-agent-failure-benchmark: beat GPT-5.4 on every axis across 6,257 traces for $1.28
- anessbelbati/jev-rerank-bench: nDCG@10 0.692 vs Cohere 0.691 (a tie, 14 datasets)
- anisselbd/jev-phishing-bench: 2,000 emails vs Claude Haiku 4.5
- iammrduncan/typesafe-ai-benchmark: Jev vs Qwen 3.8 27B on Cerebras
- vinilana/jev-eval-agent: public eval harness
- jmanhype/jev-dspy-lab: DSPy companion; calibration and confidence-gated abstention
Reproducing the interface, not the weights: proof that the decision-layer idea travels.
- TheoLeeCJ/SemIf (was "openjev"): frozen 4B model, reads option logits, runs in the browser
- vinnylarouge/jevlike: one-pass scorer; Doom, chess, and Wikispeedia demos
- Mapika/decider: Qwen3.5-2B fine-tune that emits typed decisions
- NullPo-jp/PocketJev: on-device iPhone decisions (MLX + Qwen3-VL)
Watch: the Jev Reproductions Tracker on Hugging Face follows every open-weight reproduction attempt.
The writing worth reading, ranked by writer popularity × content uniqueness. SEO recaps are excluded by design.
- Every: "TypeSafe's Jev Judged Everything I've Written in 0.7 Seconds": Mike Taylor. The only independent hands-on measurement at launch.
- TechCrunch: "A new kind of AI model from a ChatGPT inventor is thrilling developers": Mainstream validation with adoption quotes from Vercel and Bryo AI engineers.
- Flavio Copes: "A deep dive into Jev": The most thorough API/SDK walkthrough.
- Archer Hume: "Jev's Architecture Unmasked": The only serious reverse-engineering attempt.
- Latent Space: "AINews: Jev, a System One model": Swyx. The launch, placed in context.
- The Register: "TypeSafe AI debuts model for machines that plays Doom": Thomas Claburn. The best independent news coverage.
- ts2.tech: "TypeSafe AI Raises $40 Million for Jev, but Its 445× Cost Claim Is Still Self-Tested": The skeptical audit of self-tested benchmarks, plus funding verification.
- Kingy AI: "Jev review: the AI model that doesn't generate text": The skeptical counterweight.
- dev.to (Valyu AI): "How to Use Jev: A practical guide": A hands-on tutorial with code and real numbers.
- OrcaRouter: "Jev / TypeSafe System One: what we know": The claim-vs-evidence audit (the 75× vs 193× discrepancy).
- DataCamp: "System One Models: Jev": The eval-table explainer.
- agentjournal.dev: "One judge call, or twelve dimension scores?": An independent methodology measurement.
This is a community index. Each section has its own bar: see CONTRIBUTING.md for the full rules and the rubric every link is scored against. The short version:
- Official Resources: fix or update links only.
- Top Use Cases: maintainer-curated via the popularity formula; PRs may nominate but won't self-merge.
- Community Builders: open; an entry must actually use Jev, be unique, and state what it does and why it's unique.
- Benchmarks: open; label every result
[self-reported]or[independent]. - Popular Blogs: maintainer-curated via writer popularity × content uniqueness; SEO recaps are rejected.
Found something that doesn't fit any section, or just want to say hi?
- Open an issue on this repo
- Reach out on GitHub: @anandi1989
MIT: for the index itself. Individual projects retain their own licenses.
Last updated: 19 September 2026
Jev is only a few days old. Expect this list to grow fast!!!