Full Jev catalogue rebuild: direct-source research and 24-hour monitor

Run: run:manual:c338300f21a06ed237a48a75759ca44f Research checked: 25 September 2026, approximately 16:28–16:35 UTC. Monitoring window: 24 September 2026 16:27 UTC through 25 September 2026 16:35 UTC (the scheduled 16:27:15 UTC start, extended through the search). Search indexes, relative timestamps and inaccessible native search surfaces make this a bounded monitor, not exhaustive coverage. Method: official documentation and providers for mutable API/access facts; repository README plus implementation/documentation for projects; first-party model cards/repositories for alternatives. A directory, search result, self-reported benchmark, or deployed URL alone is not proof of Jev use or production operation. No account was used, no Jev call made, no app deployed, and no latency/accuracy benchmark run. All prices are displayed public listings, not account bills. Older accepted reports remain historical records, not refreshed claims.

Executive findings

1. The catalogue's two bounded negatives require revision. The six-item alternatives audit was accurate as a *22 Sep bounded search*, but today there are directly inspectable open-weight Jev-like models. TypeSafe's Jev weights were not publicly located or published in sources checked 25 September 2026; a compatible API does not make another model Jev. The old products audit was likewise a 23 Sep bounded search, not an ecosystem census. Direct Jev wiring exists in many more open projects than its four rows show, but this run still did not verify a working production customer/product deployment. Expand the project/demo section and retain the production-status bar.

2. A genuinely new, within-window project lead: Parcha AgentRun was posted in Show HN about 22 hours before the 25 Sep check. Its first-party repository documents an installable Jev adapter and an optional live-key path. The support demo is explicitly scripted and needs no key; label it *newly announced integration/SDK with scripted demo, live operation unverified*, not a live Jev product.

3. New research within the window, not product proof: Jev-Mobile was submitted 24 Sep 17:30:32 UTC; it reports Jev selecting mobile GUI actions from accessibility-tree state under VLM plans. This is an author-reported experiment; no independent replication or product deployment was checked. Jev in the Wild, submitted 24 Sep 17:46:08 UTC, studies 2,170 GitHub projects collected by 22 Sep. Its count is a paper's dataset claim, not 2,170 independently verified products.

4. Current official route: TypeSafe models still lists jev-1.13.0, aliases jev-latest and jev-preview both pointing to it, $0.042 per million direct input tokens and free output, 250,000 tokens/s and 1,200 requests/minute, 64k whole request and 32k state plus longest question. Limits are explicitly dynamic. The API is POST https://api.typesafe.ai/v1/systemone with Bearer key, model, state, and typed questions. No direct-account eligibility, live quota or credits were verified.

5. Access has expanded: add Netlify AI Gateway's official Jev route dated 17 Sep. It says Netlify Functions can use the TypeSafe SDK without separately configured API keys, billed in Netlify credits; its example is a vendor example, not a deployed customer. Vercel's model listing and launch changelog conflict on a current free/promo versus ~$0.04/M position on 25 Sep, so do not state a settled current charge without an account check.

6. X monitoring is inadequate. Public search yielded stale/irrelevant hits and no usable dated Jev post/feed for the window. The separate tools/access proposal is warranted. X's own pay-per-use page currently lists $0.005 per returned post, prepaid credits, no subscription/minimum commitment, and spending limits; the console controls current rates. A 100-post/day cap would be about $15/month in post-read charges, plus any other resource reads, but this is a planning estimate, not an account quote.

1. Current model and use-case evidence

Builder-ready use-case inventory for the use-cases page. Each row is a bounded pattern; do not present a proposed use as an existing Jev customer. E means implementation evidence in the named repository; V means vendor example; P means proposed application consistent with the API.

Retain the [unfilled workload recipe](../../../2026/09/d30a42d1-35d2-4bd7-960a-e68c1650cdb1/report.md) as a method only. No result cells may be filled from vendor or repo claims.

2. Open-source alternatives: replace the stale negative with a layered matrix

The older six rows remain useful and should be retained, but the claim “no verified open-weight Jev model found” must be dated as the 22 Sep bounded audit, not the current catalogue conclusion. The projects below are independent Jev-like implementations; none publishes TypeSafe Jev weights or proves universal parity. Code licence and weight licence are separate. Self-reported evaluation numbers are *project claims* and not a catalogue benchmark.

An additional open-alternative-jev repository was inspected as a lead. Its “one forward pass/calibrated” title is not enough to establish available trained weights or comparative quality. Do not add a parity row without checking artifacts. OpenJev contains ambitious “proven parity” numbers and named model versions that require independent reproducibility checks; not upgraded here.

3. Projects/products: expanded direct-source inventory

Status key: A = implementation shows a Jev endpoint/SDK/model path; B = first-party repository documents Jev integration but exact call file not inspected in this run; C = scripted/mocked/example; D = directory or creator lead. A/B do not establish a currently operating service, users, customers, or “production.” No row below clears the production-product bar. Direct repository URLs should be the primary links; add a live app link only with a separate access check.

Other directories JevCases, Made with Jev, awesome-jev-cobanov, awesome-jev-use-cases, and typesafeai.app are useful discovery surfaces, not row-level adoption evidence. OpenRouter's model page displays high-volume “Apps” names, but a model-page leaderboard does not establish a named product's specific feature wiring, contract, or user-facing status; do not convert it into verified product rows. Jev in the Wild provides ecosystem context, not inspection of each project.

4. Where to use, price, eligibility and terms

Conflict note: Direct TypeSafe says 64k whole request/32k state-plus-longest-question; gateway cards generally display 32k. This is a route-specific display difference, not grounds to assert the direct fan-out budget is available at every gateway. Vercel's 25 Sep price/promotion boundary is unresolved. An unauthenticated listing is not an account-approved Jev call.

5. Preceding-24-hour monitor: platform-by-platform log

All attempts took place 25 Sep approximately 16:28–16:35 UTC for the window declared above. after:/since: engine searches and relative time labels are approximate filters; exact timestamps require source metadata. Items older than the window are older newly discovered, not “new today.”

6. X tools/access proposal evidence for Manager

Recommended simplest workable option: official X API v2 recent search with a strict spend cap, if owner is willing to fund low-volume reads. X pricing says prepaid credits, no subscription/minimum, $0.005 per Post returned, $0.010/User read, $0.005/List read, daily UTC deduplication (soft guarantee), and developer-console spend limits; current rates controlled in console. X API overview describes v2 app credentials and search; pagination documents search_recent, created_at, author_id and pagination. Proposed data flow: owner creates X developer app at console.x.com, adds prepaid credit and a low monthly spending limit, supplies read-only Bearer token through secret storage, never in repo. A small scheduled script in the pool's runner environment queries recent posts at least daily (preferably every 6h to avoid missing deletions/rate gaps), saves id, URL, created_at, author, text, query, fetch time, pagination cursor and API errors as JSONL in a private dated evidence file, deduplicates IDs, then Runner follows each lead to a creator/repo/product primary source. Query terms: "jev" "typesafe", "jev-latest", "systemone" "typesafe", "@typesafeai", "jev" (github.com OR demo OR launched OR built), plus selected author accounts/lists from verified projects (TypeSafe, browser-use, Vercel Labs, Kev/CLM maintainers) after confirming handles. Search both global and known-account posts; exclude unrelated gaming “Jev” with TypeSafe/model terms. A cap of 100 new Post reads/day is ~3,000/month × $0.005 = $15/month; 20/day ~$3/month. These exclude user/list reads, taxes and any provider rate change. Start with 20–50/day and stop at budget; no purchase was made. This is a proposal, not current access.

Free fallback: owner manually exports/pastes links or a daily search digest into the pool, with post URL/time/author; Runner checks those and continues public-web search. This has low/irregular coverage and cannot establish a complete 24h feed. A self-hosted Nitter RSS/search setup is not recommended as the default: it relies on unofficial APIs and accounts, rate limits are documented in maintainer issue #1293, and its README notes 2026 takedown demands. It might cost no API fee but carries setup, account, reliability and legal/platform risk. Do not scrape public Nitter instances; its maintainer explicitly objects. Public web search is a supplementary lead source, not adequate X monitoring.

7. Publisher page-by-page change ledger

Unresolved items and next checks

• No independently verified production Jev deployment in this audit. Direct code and installable packages are real, but named customers, active workloads, current live operation, and user adoption were not substantiated. Ask maintainers for a public deployment/case study or reproducible live evidence before upgrading.

• Source access: GitHub native created/pushed search cache misses, X inaccessible as current feed, TikTok robots denial, YouTube native recent list cache miss, Reddit native new/day cache miss, DEV Go SDK repo cache miss. The platform log records negative findings as not verified, not complete absence.

• Model competition: Kev/CLM cards/repositories provide weights and code, but need independent same-workload comparisons with pinned versions, terms and realistic hardware. Laya weight/licence path and all base-model terms should be checked before publishing a licence conclusion. Avoid adopting self-reported “parity”/latency headlines.

• Route cost/terms: Vercel free-to-paid boundary, Netlify credit conversion, Cloudflare Jev dashboard rate, direct-account eligibility and all account-specific retention/credits need authenticated owner-approved checks. Recheck before any builder pilot; no real private payload in a test until terms are accepted.

• Monitoring: persist the exact daily 24h scan in PLAN.md/memory. A paid or owner-supplied X feed should supplement, not replace, primary verification. The next run should compare actual timestamps and release commits rather than treating newly discovered old repositories as new.

Historical reports are listed in Research history.