---
title: "Community repos: what people built and how they use Jev"
type: community
source_tier: community
tags: [community, repos, examples, integrations, replicas]
created: 2026-09-20
updated: 2026-09-20
confidence: medium
sources:
  - raw/x-repos/INDEX.json
  - raw/x-repos/browser-use__jev-ultrafast.md
  - raw/x-repos/awlevin__typesafe-computer-use.md
  - raw/x-repos/tamaratran__fast-jev-compaction.md
  - raw/x-repos/thruwire__foreman.md
  - raw/x-repos/realZachi__pg-jev.md
  - raw/x-repos/RomanSlack__jev-drone.md
  - raw/x/trycua-2101437979180904640.md
  - raw/x/VladTerin-2100448995415863323.md
jev_version: "jev-1.13.0"
summary: "42 third-party repos from Jev launch week, grouped by domain, each with the state it sends, the question types it asks, and the official page it maps to."
---

# Community repos: what people built and how they use Jev

> **TL;DR** Third-party projects captured 2026-09-20 from the owner's X bookmarks; stars are that day's snapshot. Each row gives the state, the primitives and what the code does with the answers. **Unaudited code** — read it before running it, and never paste a `TYPESAFE_API_KEY` into a tool you have not read. Official repos: [[entities/github-repos]].

## Start here

- **jev-ultrafast** — cleanest [[patterns/fan-out]]: operation plus a target per operation, one request.
- **typesafe-computer-use** — OCR → candidates in code → three `Choice`s + a `Noul`.
- **fast-jev-compaction** — two `Noul`s per tool call; code prunes, survivors verbatim.
- **foreman** — ten `Noul` dimensions in one call; a deterministic policy acts.
- **pg-jev** — 20 rows per state, one `Noul` per row; honest about where it degrades.
- **SemIf** — top open replica; the interface, mechanically.

## Coding agents and developer workflow

| repo | stars/lang | what it does, and how it uses Jev | maps to |
|---|---|---|---|
| [fast-jev-compaction](https://github.com/tamaratran/fast-jev-compaction) | 4918/TS | Replaces Claude Code compaction. Conversation as state → 2 `noul` per unpinned call (keep call? keep result?) → drop or truncate below 0.5. | [[cookbooks/classifying-rag-passages]] |
| [foreman](https://github.com/thruwire/foreman) | 425/Py | Supervises coding agents. Factory events → 10 `noul`s (`worker_stuck`, `work_off_track`, `needs_human`...) in one call → a deterministic policy picks STEER/STOP/VERIFY | [[patterns/confidence-routing]] |
| [jev-review](https://github.com/devagrawal09/jev-review) | 400/TS | Staged review. `Noul` risk matrix → `Choice`+`Score` file profiles → `Choice` evidence → `Choice` mechanism → `Score` severity → routing | [[patterns/composite-scoring]] |
| [jev-router](https://github.com/gargpratyush/jev-router) | 245/JS | Per-turn model routing. Prompt + tier + context tokens → four scored dimensions + confidence → fast or strong tier | [[patterns/intent-routing]] |
| [compact-adviser](https://github.com/kunchenguid/compact-adviser) | 146/TS | "Compact now?" Two one-sentence questions per request → one composed score against a floor sliding 0.90→0.50 as context fills. | [[patterns/confidence-routing]] |
| [building-with-jev-skill](https://github.com/dbreunig/building-with-jev-skill) | 124/— | Agent skill for *writing* Jev programs, not a caller: question design, state structure, thresholds; targets `jev-1.13`. | [[reference/agent-skill]] |
| [pi-warden](https://github.com/DevMortimer/pi-warden) | 102/TS | Pi guardrails. Tool call + project Markdown rules → Jev judges irreversible calls, rule breaches, stuck loops and unverified "done", then steers rather than blocks. Questions only in `docs/guards.md` | [[cookbooks/llm-guardrails]] |
| [stanley-code](https://github.com/devagrawal09/stanley-code) | 98/TS | Bounded agent workflows. One `choice` over workflow metadata + `cannot_tell`, gated on confidence ≥ 0.6, probability ≥ 0.55, margin ≥ 0.15; then per-hunk `choice` | [[patterns/intent-routing]] |
| [jev-shell-history](https://github.com/mrnugget/jev-shell-history) | 69/TS | zsh suggestions. Typed prefix + 100 ID-tagged history entries → `Choice` over IDs + `Noul` "does any complete this?" → show if top ≥ 0.3 and Noul ≥ 0.5 | [[cookbooks/semantic-find]] |
| [jev-rules](https://github.com/EliaAlberti/jev-rules) | 43/JS | Injects only the relevant standing rules. Prompt text → one `noul` per rule and map doc in one call → inject ≥ 0.6; shows all on error | [[cookbooks/skill-suggestion]] |

## Browser, computer and mobile use

| repo | stars/lang | what it does, and how it uses Jev | maps to |
|---|---|---|---|
| [jev-ultrafast](https://github.com/browser-use/jev-ultrafast) | 11208/Py | Browser agent, no screenshots. Indexed DOM element table → one request: `operation` (CLICK/TYPE_TEXT/SELECT/SCROLL/WAIT/DONE/BLOCKED) plus a speculative target question per operation. | [[patterns/fan-out]] |
| [trycua/cua](https://github.com/trycua/cua) | 25014/HTML | Computer-use platform with a `jev-use` recipe. Not in the README; per the team's article, code builds candidate actions with IDs, Jev picks one, code validates, executes and re-observes | [[concepts/state]] |
| [typesafe-computer-use](https://github.com/awlevin/typesafe-computer-use) | 582/Py | macOS computer use, ~$0.0002/step (author). OCR + accessibility tree → candidates → three `Choice`s plus an `offscreen` question per request; a `Noul` verifies the typed field. Aborts below 0.4 | [[patterns/fan-out]] |
| [mobile-jev](https://github.com/droidrun/mobile-jev) | 264/JS | Android agent, no ADB. Screen + installed apps → operation + compatible-target questions per request; code rejects stale targets. Typed text is a span from the goal; `DONE` is not proof | [[patterns/fan-out]] |
| [jev-voice-browser](https://github.com/moritzkremb/jev-voice-browser) | 148/JS | Voice to Playwright, acting on partial speech. 9–11 questions per transcript: intent/target/site/spans (`Choice`), complete/is_command/destructive (`Noul`), scroll_amount (`Score`), each thresholded. | [[patterns/confidence-routing]] |
| [unclutter](https://github.com/kitze/unclutter) | 143/TS | Hides page clutter. DOM candidates → typed keep/ad/promotion/newsletter/social/cookie/uncertain choices → hides only when probability **and** confidence are ≥ 0.9 | [[cookbooks/llm-guardrails]] |
| [WindTunnel](https://github.com/nekuda-ai/WindTunnel) | 72/HTML | WebMCP interface benchmark. Jev + Mercury 2.5 is 1 of 21 configs (49/49 with WebMCP, 25/49 with DOM controls); Jev decides, Mercury writes arguments. Runner in `experiments/jev/` | [[concepts/workflow-evals]] |

`vlad-terin/jev-browser` was announced 2026-09-17 as an adapter for Codex browser tools ("Codex plans, Jev selects, the browser moves"); it 404'd at capture, so nothing more is verified.

## Agents, orchestration and memory

| repo | stars/lang | what it does, and how it uses Jev | maps to |
|---|---|---|---|
| [skillbox](https://github.com/kitze/skillbox) | 214/TS | Self-hosted skills library over MCP. Task + skill descriptions → an **uncalibrated 0–4 rubric score** per skill (not a probability); returns ≥ 3, with a search fallback | [[cookbooks/skill-suggestion]] |
| [agentic-stack-desktop](https://github.com/codejunkie99/agentic-stack-desktop) | 64/Py | Knowledge graph across four coding agents. Not in the README; the linked post is about finding Jev-shaped decisions in a workflow | [[ideas/consult]] |
| [continual-memory-policy-model](https://github.com/codejunkie99/continual-memory-policy-model) | 5/Py | Memory policy with delayed feedback. Not in the README; its WRITE/UPDATE/DELETE/LINK/COMPACT/NOOP action set is Jev-shaped, but it trains its own policy | [[concepts/primitives]] |
| [octavus-dual-brain](https://github.com/leotavares26/octavus-dual-brain) | 0/HTML | Write-up placing Jev as System 1 ahead of an orchestrator. `intent` (`Choice`) + `risk_score` (`Score` 0–3) in parallel; 10 filings fanned out in one call | [[patterns/intent-routing]] |

## Search, data and databases

| repo | stars/lang | what it does, and how it uses Jev | maps to |
|---|---|---|---|
| [jev-search](https://github.com/superagents-lab/jev-search) | 286/TS | Plain-language web search, no generated answers. Typed questions choose query, sources and time range; Jev scores each result, code orders by relevance and engine agreement | [[cookbooks/rerank]] |
| [pg-jev](https://github.com/realZachi/pg-jev) | 239/Shell | Postgres extension: `jev(table,'condition')` in `WHERE`. 20 rows per request in one shared state, one `Noul` per row. Warns positional `rows[i]` lookup degrades in long arrays | [[cookbooks/parallel-questions]] |
| [neo4jev](https://github.com/jexp/neo4jev) | 46/Jup | Graph navigation one hop at a time. Outgoing relationships as `Choice` options + a `Noul` "goal reached?" in the same call → top-k over probabilities is a beam search | [[cookbooks/hierarchical-classification]] |

## Moderation, judging and media

| repo | stars/lang | what it does, and how it uses Jev | maps to |
|---|---|---|---|
| [notra](https://github.com/usenotra/notra) | 193/TS | GEO platform tracking brand mentions. Not documented — no Jev or TypeSafe mention in the README | — |
| [youtube-sponsor-detection](https://github.com/trungdq88/youtube-sponsor-detection) | 78/JS | Skips YouTube sponsor reads. Transcript as numbered lines → Jev picks line IDs (`Choice`) plus probability questions, each repeating the sponsor definition. Code owns every timestamp; skips above ~70% | [[cookbooks/semantic-find]] |
| [jevmeter](https://github.com/ChetasLua/jevmeter) | 75/Py | Live "BS meter" over video. One request per sentence; state = sentence + speaker history + last question → 5 preset `noul`s → averaged into a gauge, flagged at the 98th percentile | [[cookbooks/consistency-noul]] |
| [Jev-Moderation-Bot](https://github.com/brainstormity/Jev-Moderation-Bot) | 39/Py | Discord spam/scam filter. Messages + prior infractions → five dimensions (scam, spam, noobness, toxicity, helpfulness) → archetype plus two admin tiers; pardons fed back | [[patterns/composite-scoring]] |
| [jev-as-a-judge](https://github.com/danielgshea/jev-as-a-judge) | 25/Py | Jev vs three LLM judges on fixed runs. Trace as state → `does_pass` (`Noul`) + a `quality` score. Reports 100% oracle agreement over 500 repeats, variance 92–913× below the LLM judges | [[guides/testing-and-evaluation]] |

## Real-time, games, robotics and IoT

| repo | stars/lang | what it does, and how it uses Jev | maps to |
|---|---|---|---|
| [jev-trader](https://github.com/jarrodwatts/jev-trader) | 1464/TS | One decision per Monad block. Order-book snapshot → buy/sell `Choice` per ~300 ms; code posts the order. **Numeric time-series is a documented Jev weakness** — a latency demo, not a trading system | [[concepts/jaggedness-jev-1-13]] |
| [typesafe-mario](https://github.com/fhshaik/typesafe-mario) | 295/Py | Plays Super Mario Bros from emulator RAM. JSON telemetry → `Choice` picks the controller macro, `Noul` whether a jump helps, `Score` immediate danger. No scripted override | [[guides/choosing-a-primitive]] |
| [jevpilot](https://github.com/standardagents/jevpilot) | 112/JS | Three.js driving-sim autopilot. Compact tables of eligible paths, boundaries, traffic, signals and stop memory → a choice over paths; geometry stays local | [[concepts/state]] |
| [jev-drone](https://github.com/RomanSlack/jev-drone) | 83/Py | Camera-only quadrotor in MuJoCo, ~2.5 Hz. Per call: `maneuver` (`Choice`), `risk` (`Score`), `target_truly_lost` (`Noul`) → search gates on the **Noul**, not the unconfident Choice. Jev never picked `climb` until the state named altitude | [[guides/writing-instructions-and-criteria]] |
| [HA-Jev](https://github.com/AboveColin/HA-Jev) | 32/Py | Home Assistant: questions become sensors. Pickers build the state → `jev.noul`, `jev.choice` (2–255 options), `jev.score` (2–10 levels), `jev.ask`; one context is one request | [[guides/smart-home-demo]] |

## Integrations and catalogs

| repo | stars/lang | what it does, and how it uses Jev | maps to |
|---|---|---|---|
| [awesome-jev-by-typesafe](https://github.com/Anil-matcha/awesome-jev-by-typesafe) | 684/Py | Curated use cases and starter code; a directory, not a caller. Quick start matches SDK 0.6.0; pattern list mirrors the official four | [[patterns/overview]] |
| [typesafe-mcp](https://github.com/itsmostafa/typesafe-mcp) | 133/Go | MCP server with one `evaluate` tool. Passes `state` + a `questions` map (`{type, instructions, criteria?}`) to `POST /v1/systemone`, default `jev-latest`; validates criteria locally | [[reference/http-api]] |

## Open replicas and Jev-compatible servers

None affiliated with TypeSafe. Numbers compared in [[ideas/field-reports]].

| repo | stars/lang | what it is, and its relation to Jev | maps to |
|---|---|---|---|
| [SemIf](https://github.com/TheoLeeCJ/SemIf) | 2310/Py | Semantic ifs from open models on a 3090. Reads option logits from a frozen 4B model in one forward pass; ~1.0 s for 21 probability pairs, conditional on the supplied options | [[syntheses/jev-vs-llm-structured-outputs]] |
| [NanoJev](https://github.com/TianyuCodings/NanoJev) | 1322/Py | 0.6B parallel decision model + training pipeline. Same request shape (state, question, 2–255 candidates); Choice via set attention, Boolean via sigmoid, Score as a weighted level. | [[concepts/primitives]] |
| [jevlike](https://github.com/vinnylarouge/jevlike) | 1059/Py | Starter model with Jev's input/output shape. Option-attention head, softmax across options; ~100× faster than a decoder writing 400 tokens. Reused by trycua CUA-S1-FORMS | [[syntheses/jev-vs-llm-structured-outputs]] |
| [openjev-sglang](https://github.com/ekzhang/openjev-sglang) | 214/Py | Jev-compatible `/v1/systemone` on Qwen3.6-35B. N+1 one-token calls over a shared prefix; Noul = P(yes), Choice = argmax + distribution, Score = `sum(level_index * probability)`. Its `confidence` is reimplemented | [[concepts/confidence]] |
| [mini-jev](https://github.com/r-ms/mini-jev) | 29/Py | Letter-logit reads vs grammar-constrained JSON on Qwen3-4B. 4× faster on short texts. Making a model *write* probabilities scored 0.346 vs 0.896 for letter reads; shares are not calibrated | [[concepts/confidence]] |
| [openjev](https://github.com/TheoLeeCJ/openjev) | 2310/Py | Same URL, stars and README as SemIf: one project after a rename ("SemIf (formerly OpenJev)"). Count it once | see SemIf |

## Reusable designs worth stealing

- **Speculative target heads** — ask the operation *and* a target question per operation in one request; run only the matching target. *jev-ultrafast, mobile-jev* → [[patterns/fan-out]].
- **Score every context block for keep/drop** — two `noul`s per item, prune in code, survivors verbatim. *fast-jev-compaction* → [[cookbooks/classifying-rag-passages]].
- **Model assesses, code decides** — many `noul` dimensions per call, then a deterministic state machine acts. *foreman* → [[patterns/confidence-routing]].
- **Staged escalation** — each stage's strongest signal selects the next question. *jev-review, stanley-code* → [[patterns/intent-routing]].
- **Rule or skill selection per prompt** — one yes/no question per candidate, one request, inject above a threshold. *jev-rules, skillbox* → [[cookbooks/skill-suggestion]].
- **Split one big action choice into narrow ones** — overlapping options read as low confidence. *typesafe-computer-use* → [[guides/writing-instructions-and-criteria]].
- **Select the span, never generate it** — regex or numbered lines make candidates, Jev picks one, code copies it. *jev-voice-browser, sponsor-detection* → [[cookbooks/pre-parsed-value-extraction]].
- **Batch N items into one state with N questions** — the shared preamble amortises state tokens; positional lookup into long arrays is the failure mode. *pg-jev* → [[cookbooks/parallel-questions]].
- **Gate on the primitive that fits** — when a `Choice` is unconfident, a companion `Noul` may answer cleanly. *jev-drone* → [[guides/choosing-a-primitive]].

## Caveats

- **Stars are a launch-week snapshot** (2026-09-20); several repos were days old. Hype signal, not quality.
- **Licences missing** on Jev-Moderation-Bot, jev-as-a-judge, building-with-jev-skill, openjev-sglang, typesafe-mario, jev-shell-history, jevpilot and sponsor-detection; pg-jev is `NOASSERTION`.
- **One duplicate**: `TheoLeeCJ/openjev` and `TheoLeeCJ/SemIf` are one repository after a rename — identical URL, stars and README.
- **Jev usage not in the README** for notra (no mention at all), agentic-stack-desktop, continual-memory-policy-model, trycua/cua (only its X article) and WindTunnel. pi-warden and Jev-Moderation-Bot name the judgments, not the types.
- **Community limit claims**: fast-jev-compaction calls 32k "Jev's request ceiling"; compact-adviser caps requests at 32,000 bytes. Documented rule: 64k tokens per request, 32k for `state` plus the **longest single question** ([[reference/models-and-pricing]]) — those are approximations.
- **Uncalibrated numbers**: skillbox's 0–4 rubric is not a probability, and openjev-sglang computes its own entropy-based confidence ([[concepts/confidence]]).
- **Security**: unaudited code, much of it agent plugins with shell, browser or database reach. Several ship smoke scripts that bill your key. Never pass a key as a CLI arg.

## Related

- [[ideas/consult]] — answering "could Jev help this project?"
- [[ideas/patterns]] — the patterns these repos share, with fit verdicts
- [[ideas/field-reports]] — independent evaluations and replica numbers
- [[entities/github-repos]] — the official `typesafe-ai` repos

## Sources

- raw/x-repos/INDEX.json and the 42 per-repo captures there (metadata + README, fetched 2026-09-20)
- raw/x/trycua-2101437979180904640.md — only source for trycua/cua's Jev usage
- raw/x/VladTerin-2100448995415863323.md — only source for vlad-terin/jev-browser