Jev – a curation of Jev demos on X, tools, skills, and integrations

以下正文同步自 Hacker News，版权归原站所有，已转换为易读排版。

From an interesting demo to code you can actually understand.

A curated starting point for TypeSafe Jev: reviewed projects, reusable skills, small code examples, and creator demos from X, GitHub, and YouTube.

33 reviewed picks — start with the source-reviewed selection, not the raw index.

Browse reviewed picks · Freshness and link health

Evidence legend: Reviewed = primary source inspected · Indexed = community listing · Auto-discovered = README keyword match. These labels are not security certifications.

Last discovery: 2026-09-22T13:30:23Z (UTC). Latest repository metadata date: 2026-09-22. Unavailable repositories at last check: 13. Coverage, failures and run receipts.

Navigate: Quick start · Choose a path · CUA & drivers · Reviewed picks · OpenRouter winners · Watch demos · Search the catalog · Contribute

Source-reviewed does not mean security-audited. Keep credentials private and inspect permissions before running a project.

Jev is TypeSafe AI's model for answering typed questions about text or JSON: choose an option, estimate whether something is true, or score it against a rubric. Use it for bounded decisions such as routing, ranking, or selecting an agent's next action; use a generative LLM when you need new prose, code, or an open-ended plan. Your application still controls permissions, validation, and execution. Official documentation · Official Python quick-start source.

JSON in → one typed decision out. With typesafe-sdk installed and TYPESAFE_API_KEY set, this makes one billable API request:

from typesafe_sdk import Choice, TypeSafeClient

with TypeSafeClient() as client:

result = client.system_one(

state={"document": "I was charged twice."},

questions={"team": Choice(

instructions="Which team should handle this ticket?",

criteria={"billing": None, "technical": None, "other": None},

)},

)

print(result.choices["team"].choice) # e.g. billing; actual output can differ

No key yet? Clone this repository and run python3 examples/route_ticket.py to inspect an offline request without making an API call. Setup and next steps · More question packs.

Browser Use / @gregpr07. Author recording, not our benchmark. Implementation · More demos and videos.

Build your first typed classifier → JSON input, a Choice question, and your first response.

Use Jev in an agent → Select observed actions, validate them, and keep a fallback.

Explore a local alternative → Independent implementations, hardware requirements, and important differences.

Computer-use map: integrations, drivers, skills, and local alternatives.

Native/desktop: Cua Driver + jev-use, Jev Cua, and jev-bot. Browser: compare Jev Ultrafast, the smaller browser skill, Jev Browser, and the existing-connection Codex skill in the map.

Keep the layers separate: Jev selects; the driver executes; application code verifies. Browser Harness and Cua Driver are supporting runtimes, not Jev models. CUA-S1-FORMS and Jevlike are independent local-model projects, not official Jev weights. A locally installed agent may still call hosted Jev.

These are the editorial selections. Follow a project immediately, or open the full reviewed page for its concrete Jev role, source evidence, review date, and reuse notes.

Start building: TypeSafe JavaScript SDK · TypeSafe Python SDK · Jev Logs · Jevify · Official TypeSafe skill · Skillbox.

Use it in an agent or app: Cua Driver + jev-use · Jev Browser (MCP, CLI, library) · Jev Browser Skill · Jev Browser Use (existing Codex connection) · Jev Cua (Codex/Waku) · Jev Ultrafast · Jev voice browser · jev-bot (@compootor) · TypeSafe computer use (macOS) · Foreman · Jev MCP · jev_search · TypeSafe MCP · tisco · Vibe Domain · Browser Harness (supporting driver) · Home Assistant Jev.

Explore games and creative tools: Jev Chess · Jev Plays Snake · JevAI for XMage · JevPilot · PROMPT FPS / Jev Doom · TypeSafe Mario · Jev Music Playground.

Investigate local alternatives: CUA-S1-FORMS (independent specialist) · Jevlike (independent option scorer) · OpenJev SGLang.

All 33 reviewed picks — with evidence, dates and reuse notes

OpenRouter's 21 September 2026 showcase named five projects: JevAI for XMage, Jev Chess, tisco, Vibe Domain, and jev_search.

Explore game decisions, transcript search, domain ranking, and two-pass code search. Each winner has its original post and publisher screenshot in the visual gallery, plus source and implementation notes. Three have linked source repositories; the other two are public project pages, not verified open-source releases. “Winner” is OpenRouter's designation, not our ranking or a security certification.

On GitHub: open the categorized catalog, choose a heading from its outline, and use Cmd/Ctrl+F for a project, creator, or keyword.

Prefer filters? The local catalog viewer starts with reviewed picks. Search by keyword, then filter by evidence level, category, available code/skill/demo, and license. It runs without dependencies, sign-in, API keys, or a server; the viewer's data is embedded at build time. Browse instructions · Machine-readable catalog.

A compelling clip often leaves the useful questions unanswered: what did Jev decide, where is the implementation, and what can I reuse? This directory connects the original demo to the code and separates inspected sources from the much larger discovery queue. A high raw entry count is not the goal; a shorter path to something useful is.

A concrete Jev integration, demo, skill, SDK, reusable snippet, tutorial, or clearly labeled independent reproduction—with a traceable primary source and a specific explanation of Jev's role. Generic type-safety libraries, unrelated projects named Jev, unsupported performance claims, and duplicate listings do not belong.

Repositories are deduplicated by canonical owner/name; renamed projects should point to their successor. Maintainers can correct an entry, remove it from the reviewed selection, or exclude it from future imports. Inclusion, deduplication, and removal policy.

Suggest a demo or correction with the original creator, a source link, what Jev decides, and any code, skill, or recording. For a PR, edit the relevant JSON source and run python3 scripts/build.py, then python3 -m unittest discover -s tests -v and python3 scripts/build.py --check. Contributor guide.

The four-hour workflow performs bounded discovery and rotating checks; it does not continuously inspect every website. Freshness, unavailable repositories, media errors, and run receipts distinguish actual checks from scheduled intentions.

Source review is not a security audit, successful execution test, benchmark replication, or endorsement. Inspect permissions and outbound data before installing hooks, browser extensions, or agent skills. Keep API credentials private; demos can use paid inference, privileged desktop access, or logged-in browser profiles.

Recordings, illustrations, fixtures, and mock modes are labeled where known. A live-looking dashboard is not evidence of real inference or profitable trading. Independent reproductions are not official Jev weights or proof of equivalent quality.

Missing licenses remain not-checked, not-detected, or NOASSERTION, not assumed permission to reuse. Original code and writing follow the repository's license; imported metadata and linked media retain their original terms. This is an unofficial community directory, not endorsed by TypeSafe AI. Security policy · Attribution.

Original creator media, linked to the implementation. Click a preview to watch the source. Animated GIFs play inline; original GitHub video attachments appear as players below. X and YouTube open on their own platforms.

Browser Use × Jev Ultrafast

Watch a browser choose operations and DOM targets; a separate model writes text.

Browser Use · @gregpr07 — Animated author recording. This narrow flight-search demonstration is not a general reliability benchmark.

Partial speech becomes typed intents and observed page targets; Playwright acts.

Moritz Kremb · @moritzkremb — Creator post preview when accessible. Keep its browser-control server private; a recording is not a security audit.

OCR and accessibility observations become choices for desktop actions.

@awlevin — Creator post preview when accessible. Screen content may be private; real use needs screen-recording and accessibility permissions.

Jev assesses worker progress and verification needs; code decides interventions.

Josh Rosen · @JoshARosen — Architectural experiment. Not a replacement for tests, review, or permission checks.

No reliable preview image was returned. The original watch link is retained.

Typed buy/sell judgments inside a bounded quoting and execution loop.

Jarrod Watts · @jarrodwatts — The documented default is mock/dry-run. A dashboard does not prove live inference, profitability, or safe custody.

The discovery thread that seeded this collection; follow the original creators.

Moritz Kremb · @moritzkremb — Roundup image, not a separate working project. This directory does not claim to capture every reply.

Legal game actions become Jev choices; the hybrid combines XMage search with model judgment.

ShiftSad · screenshot published by OpenRouter — Publisher screenshot, not a recording or our test. The reported 11-6-3 result belongs to the hybrid; pruning is off by default. Award announced 2026-09-21.

A shared chess game displays Jev probabilities over legal move candidates.

sliday · screenshot published by OpenRouter — Publisher screenshot; Source opens the project page. No public repository or code license verified. We did not play a move or reproduce the results. Award announced 2026-09-21.

Search spoken content, review transcript matches, and approve clip organization in a terminal.

cairodavila · screenshot published by OpenRouter — Publisher screenshot. Jev judges transcripts, not video pixels. Audio uploads, live model calls, and file changes have separate data and approval implications. Award announced 2026-09-21.

Separate candidate availability checks from ranking names against a requested vibe.

OB Studio / Oliver · screenshot published by OpenRouter — Publisher screenshot; Source opens the project. The page documents a no-key heuristic mode distinct from live Jev. No public repo verified; recheck availability at a registrar. Award announced 2026-09-21.

Keyword-density code prioritizes files; Jev evaluates passages in two passes.

caio0452 · screenshot published by OpenRouter — Publisher screenshot. The author labels the project AI-generated and not for production. Selected source text reaches OpenRouter; no project code was executed here. Award announced 2026-09-21.

Jev Tetris: see the decision

Code enumerates reachable placements; Jev chooses and the board visualizes its preferences.

thelau — Author UI still. A mock mode also exists; the screenshot alone does not establish live inference.

Follow staged risk checks, evidence selection, and severity judgments in a local UI.

devagrawal09 — Author dashboard screenshot. Findings are review prompts, not proof of a defect or a security certification.

Score source passages and select the relevant sentence without generating a new answer.

Shubhamsaboo / awesome-llm-apps — Author screenshot of the semantic-find interface. Live requests use Vercel AI Gateway and send selected text.

A prompt becomes a musical plan, bounded choices, editable notes, and MIDI.

cocktailpeanut — Author architecture illustration, not a screenshot or an audio-model result. The full recording appears below.

Explore Noul, Choice, and Score through an interactive decision playground.

davila7 — Explanatory illustration, not an actual response or timing result. Hosted live examples require a key.

Local search proposes files and history items; Jev ranks the intended targets.

dabit3 — Historical author screenshot from the archived experiment. Its README points to a successor; this is not the current UI.

Jev ranks a bounded set of previous commands; the user a

正文由 FLUX 从来源站点 RSS 同步，内容未经改写；遇到排版缺失或需要图片、视频时请以原文为准。