~/wiki/ideas
Community repos: what people built and how they use Jev
[ community tier ][ updated 2026-09-21 ][ confidence medium ][ jev-1.13.0 ]#community · repos · examples · integrations · replicas
TL;DR 58 third-party projects from Jev's first week (captured 2026-09-20; 16 more 2026-09-21, stars from those days). Read this index, then one domain page — Repos: coding agents, orchestration, memory, browser and computer use, integrations or Repos: data, documents, judging, real-time, markets, business apps, replicas — which carry the links and per-repo detail. Unaudited code: vet it (below). Official: typesafe-ai GitHub organisation and repos.
Start here
- jev-ultrafast — cleanest Speculative fan-out: operation plus a target per operation, one request.
- fast-jev-compaction / winnow — a
noul per item, code prunes, survivors verbatim; winnow adds a replay harness.
- 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.
- docjev — best-documented measurement: 40 real PDFs, timings, costs.
- SemIf — top open replica; the interface, mechanically.
All 58 repos
Domains 1–4 → Repos: coding agents, orchestration, memory, browser and computer use, integrations; 5–11 → Repos: data, documents, judging, real-time, markets, business apps, replicas.
1. Coding agents and dev workflow
| repo |
decides |
| tamaratran/fast-jev-compaction |
keep or drop each tool call |
| GhalebDweikat/winnow |
keep or hide each block of a tool result |
| thruwire/foreman |
worker stuck, off-track, needs a human? |
| qkal/Canny |
does this message claim "done"; rule broken? |
| devagrawal09/jev-review |
risk, evidence, mechanism, severity per diff |
| devagrawal09/stanley-code |
which bounded workflow this is |
| 0xNatoshi/jev-codex-router |
model tier and thinking depth per Codex call |
| gargpratyush/jev-router |
fast or strong tier for this turn |
| kunchenguid/compact-adviser |
compact now, against a sliding floor |
| DevMortimer/pi-warden |
irreversible, off-task or looping call? |
| EliaAlberti/jev-rules |
which standing rules this prompt needs |
| mrnugget/jev-shell-history |
which past command completes this prefix |
| ellipsis-dev/blink |
which file or folder the query is in |
| dbreunig/building-with-jev-skill |
nothing — teaches agents to write Jev |
| ryana/jevify |
nothing — a prompt that audits a repo for Jev-shaped work |
2. Browser, computer and mobile use
| repo |
decides |
| browser-use/jev-ultrafast |
which browser operation, on which element |
| trycua/cua |
which prebuilt computer action next |
| awlevin/typesafe-computer-use |
which macOS element; did typing land |
| droidrun/mobile-jev |
which Android operation, which target |
| moritzkremb/jev-voice-browser |
what the half-spoken command meant |
| kitze/unclutter |
is this DOM node clutter |
| nekuda-ai/WindTunnel |
benchmark: Jev picks the WebMCP tool |
3. Agents, orchestration and memory
| repo |
decides |
| kitze/skillbox |
which skills a task needs (0–4 rubric) |
| codejunkie99/agentic-stack-desktop |
not in the README |
| codejunkie99/continual-memory-policy-model |
write/update/delete/link/compact |
| leotavares26/octavus-dual-brain |
intent and risk before an orchestrator |
| kerpopule/hermes-jev-skills |
model, memory, compaction, skill and GUI action per turn |
4. Integrations and catalogs
| repo |
decides |
| jkudish/jev-mcp |
ten typed judgment tools over MCP |
| itsmostafa/typesafe-mcp |
one evaluate tool over MCP |
| sharziki/semdecide |
typed decisions as Unix exit codes |
| Anil-matcha/awesome-jev-by-typesafe |
nothing — a catalogue |
| vercel-labs/json-render |
not Jev-specific; one experimental path |
5. Search, data and databases
| repo |
decides |
| superagents-lab/jev-search |
which result answers the query |
| realZachi/pg-jev |
a semantic predicate in SQL WHERE |
| jexp/neo4jev |
which graph edge to walk next |
| AkashPriyadarshii/jev-curate |
keep or drop each training-data row |
6. Documents
| repo |
decides |
| jerryjliu/docjev |
a document's category; where a packet splits |
| kyotofin/tax-doc-classifier |
which IRS form a page is |
7. Moderation, judging and media
| repo |
decides |
| trungdq88/youtube-sponsor-detection |
which transcript lines are sponsor reads |
| ChetasLua/jevmeter |
a per-sentence "BS meter" over video |
| brainstormity/Jev-Moderation-Bot |
scam, spam, toxicity per message |
| danielgshea/jev-as-a-judge |
did this captured agent run pass |
8. Real-time, games, robotics and IoT
| repo |
decides |
| fhshaik/typesafe-mario |
which controller macro this frame |
| standardagents/jevpilot |
which driving path to take |
| RomanSlack/jev-drone |
maneuver, risk, is the target lost |
| AboveColin/HA-Jev |
house questions as HA sensors |
| emrickgarrett/OneVOneJev |
an FPS opponent's move, aim, fire, ~9 Hz |
9. Markets — engineering only
| repo |
decides |
| jarrodwatts/jev-trader |
buy or sell per Monad block; latency demo |
10. Business and idea tools
| repo |
decides |
| monteduro/killmyidea |
eight 0–4 scores → KILL / FIX / SHIP |
| usenotra/notra |
not in the README |
11. Open replicas and Jev-compatible servers
| repo |
decides |
| TheoLeeCJ/SemIf |
option logits from a frozen 4B model |
| TianyuCodings/NanoJev |
0.6B trained replica plus pipeline |
| vinnylarouge/jevlike |
starter model with Jev's I/O shape |
| ekzhang/openjev-sglang |
Jev-compatible /v1/systemone |
| r-ms/mini-jev |
letter-logit reads vs constrained JSON |
| TheoLeeCJ/openjev |
the same repo as SemIf, renamed |
Linked but not about Jev. lahfir/agent-desktop (1399★, Rust) and irfndi/prism-liquidity-agent (59★, TS) name neither Jev nor TypeSafe in their READMEs; they are here only because @charliejhills' roundup links them. vercel-labs/json-render carries the same caveat.
Reusable designs worth stealing
- Speculative target heads — the operation and a target question per operation. jev-ultrafast → Speculative fan-out.
- Score every block keep/drop — a
noul per item, prune in code, survivors verbatim. fast-jev-compaction, winnow → Cookbook: Classifying RAG passages.
- Model assesses, code decides — many
noul dimensions per call, then a deterministic state machine acts. foreman, Canny → Confidence-gated routing.
- Staged escalation — each stage's strongest signal picks the next question. jev-review → Intent routing.
- One yes/no per candidate rule or skill, one request, inject above a threshold. jev-rules, skillbox → Cookbook: Skill suggestion.
- Split one big action choice into narrow ones — overlapping options read as low confidence. typesafe-computer-use → Writing instructions and criteria that Jev reads correctly.
- Select the span, never generate it — regex or numbered lines make candidates. jev-voice-browser, jev-mcp → Cookbook: Pre-parsed value extraction.
- Batch N items into one state, N questions — positional lookup into long arrays is the failure mode. pg-jev → Cookbook: Parallel questions.
- Gate on the primitive that fits — an unconfident
Choice, a clean companion Noul. jev-drone → Choosing between Choice, Score, Noul.
- Validate each answer, fail closed per item — in-set, maximal, sums to 1. jev-mcp → HTTP API: POST /v1/systemone and GET /v1/models.
- Ship a replay harness with the thresholds — score the judge offline on your own history. winnow, Canny → Testing and evaluating a Jev workflow.
Caveats
- Vet before you install (@k2sbhai): most were written in launch week, and several send prompts, diffs, tool inputs or on-screen text to TypeSafe. Read the source; check what leaves the machine; dedicated API key; dry-run or shadow mode first; set confidence thresholds; cap max actions.
- Stars are a launch-week snapshot and have moved since: @socialwithaayan put jev-ultrafast at 15k on 2026-09-21 against the 11,208 in our table. Several repos were days old. Hype, not quality.
- No licence on Jev-Moderation-Bot, jev-as-a-judge, building-with-jev-skill, jevify, openjev-sglang, typesafe-mario, jev-shell-history, jevpilot, sponsor-detection, blink, OneVOneJev, killmyidea; pg-jev is
NOASSERTION. openjev = SemIf, renamed.
- Jev usage not in the README for notra, agentic-stack-desktop, continual-memory-policy-model, trycua/cua (X article only) and WindTunnel; pi-warden and Jev-Moderation-Bot name judgments, not types.
- Community limit claims are local caps, not Jev's: 32k "request ceiling" (fast-jev-compaction), 32,000 bytes (compact-adviser), 120k chars / 40 questions (jev-codex-router), 120,000 chars / 200 blocks (winnow), 1,000,000 bytes (semdecide). Documented: 64k tokens per request, 32k for
state plus the longest question (Models, aliases, pricing, rate limits, context).
- Uncalibrated or unverified: skillbox's 0–4 rubric is not a probability; openjev-sglang computes its own confidence (Confidence vs probability); Canny asserts ~0.05 run-to-run drift, jev-curate 1,500 rows/sec and cross-vendor cost ratios; jev-mcp's OpenRouter, Cloudflare and Vercel routes are in no first-party doc.
- Security: unaudited code, much of it plugins with shell, browser, editor or database reach. Several ship smoke scripts that bill your key. Never pass a key as a CLI arg.
Related
- Repos: coding agents, orchestration, memory, browser and computer use, integrations, Repos: data, documents, judging, real-time, markets, business apps, replicas — every repo above in full, with links
- Consult guide: could Jev help this project? — "could Jev help this project?"; Decision patterns from the community (with fit verdicts) — shared patterns, fit verdicts
- Field reports: independent evaluations, critiques, open replicas — independent evaluations; typesafe-ai GitHub organisation and repos — official repos
Sources
- raw/x-repos/INDEX.json and the 58 per-repo captures beside it (42 on 2026-09-20, 16 on 2026-09-21); each domain page lists the captures it covers
- raw/x/charliejhills-2102044071368012140.md and raw/x/socialwithaayan-2102059089652285739.md (roundups), raw/x/k2sbhai-2101657436696547773.md (vetting checklist)
