把读屏+意图判断+候选回复做成手机上的对话副驾,解决微信飞书里不会回话的痛点,不注入不封包降低了封号顾虑;1.5天3723星来自极强话题性。商业化可走订阅制输入法式入口。最大风险是平台ToS与隐私合规,且同族仓库刷量嫌疑会反噬可信度。
装在手机上的对话副驾:在微信 / QQ / X / 飞书里读懂对方、给出候选回复、一键填入输入框,发不发由你。非侵入,只读屏幕,不 hook 不改包。
★3723
Hidden gems, rated before they trend.
今天发现 320 个新项目,174 个评了级。
收录新建 2026-09-21 – 2026-09-23抓取 600粗筛 50已评级 41
把读屏+意图判断+候选回复做成手机上的对话副驾,解决微信飞书里不会回话的痛点,不注入不封包降低了封号顾虑;1.5天3723星来自极强话题性。商业化可走订阅制输入法式入口。最大风险是平台ToS与隐私合规,且同族仓库刷量嫌疑会反噬可信度。
装在手机上的对话副驾:在微信 / QQ / X / 飞书里读懂对方、给出候选回复、一键填入输入框,发不发由你。非侵入,只读屏幕,不 hook 不改包。
★3723
0.6B的System 1决策模型,把diff、trace、日志这类状态一次前向映射为校准概率,50ms出结果、零解码token,省的是agent推理成本;1.3天202星,是同族中最接近模型本体的资产,可做agent标准依赖。风险是同类命名混乱、指标缺第三方验证。
AgentJev-0.6B - a fast 'System One' decision model for AI Agents: feed it any unstructured state (diffs, traces, logs) and structured questions, get calibrated probability distributions back in one ~5…
★202
把 Qwen-Image 搬到 ncnn + Vulkan 上跑,作者是 waifu2x-ncnn-vulkan 的 nihui,无独显本地出图是真需求。仓库暂无描述,需读代码确认支持的模型与显存占用;可对照 stable-diffusion.cpp 一类本地推理方案。
★42
借 Windows 子会话让第二桌面独立跑游戏,再用 Sunshine 串流,主桌面照常使用,思路少见且切入点非常具体。对单机多开或远程串流的人是实招;限制是依赖系统特性,稳定性和反作弊兼容需自测。
Stream games from a second Windows desktop while you keep using your PC - via Windows child sessions + Sunshine
★43
在Apple silicon上做本地computer use,用Accessibility API填表加本地小模型做类型化决策,解决云端CUA的隐私与延迟问题;131星两天说明本地隐私路线有拥趸。可卖企业内网自动化。风险是苹果权限收紧,以及官方computer use与Apple自身能力的正面碾压。
Local computer use on Apple silicon: on-device typed decisions with laya and form filling with CUA-S1-FORMS, driven through the macOS Accessibility API
★131
在 AI agent 与 MCP server 之间加一道零信任网关:能力白名单、脱敏、速率与成本上限、可审计会话日志。MCP 铺开后权限失控是真问题,这类网关还很少;刚建一天,能否真防住越权要看实现。
Zero-trust security gateway between AI agents and MCP servers: scoped capabilities, redaction, rate and cost limits, and verifiable session logs.
★20
自托管的社媒排期工具,一次写好分发到多个平台,跑在自己的 Cloudflare 账号上,小团队和个人内容号今天就能部署使用。思路不新,对标 Buffer 一类付费服务;胜在范围克制、说明清楚,是能上手的那类项目。
Self-hosted social media scheduler. Write once, publish to any connected platform. Runs on your own Cloudflare account.
★63
用 Shell 在 Apple Silicon + Parallels 里自动装 Arch Linux ARM,正好绕开官方 ISO 只有 x86_64 这个具体坑,省掉一晚上折腾。范围克制;局限是只覆盖 Parallels 一种虚拟化。
Mac 虚拟机安装 Arch Linux 懒人包|在 Apple Silicon + Parallels Desktop 上自动安装 Arch Linux ARM(官方 ISO 仅 x86_64,无法直接启动)
★28
针对开源音乐生成模型 YuE2 的并发推理优化,给出每首歌 1.68×、单卡曲目数 3.31× 的具体数字,模型与配方不变。自己跑音乐生成的人直接可用;局限是只服务 YuE2 一个模型。
Fast, concurrent inference for YuE2. Same model and recipe: 1.68× faster per song, 3.31× more songs per GPU.
★47
把特斯拉车机系统跑进 QEMU,作者是车联网安全圈的熟面孔,对做车载固件逆向和车机研究的人是少见且可复现的入口。描述只有一行,缺步骤与镜像获取说明;同类的多是零散博客,这样成型的仓库不多。
Run Tesla OS On QEMU
★52
Rails 的多租户库,来自 GoRails 作者,只解决一件事且描述一句话讲清,装上就能用。概念不新,acts_as_tenant 等已有成熟方案;值得看的是它如何用更小的接口处理租户隔离。
Multitenancy for Rails
★25
一个让 Agent 用 Three.js 生成可交互、能动态自组装 3D 建筑的技能包,把生成式编码落到具体视觉产物上,做网页 3D 演示的人能直接试。星速很快,但能否稳定产出好效果还得看示例质量。
An Agent Skill for building interactive, self-assembling 3D architecture with Three.js. 一个用 Three.js 生成可交互、可动态建造 3D 建筑的 Agent Skill。
★101
yoheinakajima 的新作:在本机跑开源视觉模型,对图片提有类型的问题并拿回概率而不是一段话。思路清爽、边界明确,适合要拿 VLM 做判定的人;限制是仅限本地小模型,精度取决于所选权重。
Ask an open vision-language model typed questions about an image and get probabilities back, on your own machine.
★13
给 Google Antigravity 做多账号并行与沙箱隔离的原生工具,针对新编辑器额度紧张的人,动手很快。切入点及时、需求真实;但依赖未公开的内部机制,官方一改就可能失效,长期可用性存疑。
Native, non-invasive parallel multi-account & sandbox manager for Google Antigravity
★43
把路由、判断与打分做成 0.4B 双向模型的单次前向,输出校准置信度;对想在 Agent 里替掉生成式判定的团队有用。描述克制、架构具体,但没说明权重是否放出,同题材近两天扎堆出现,值得先看再定。
Dev (dev-0.4b): a 0.4B bidirectional decision model on ModernBERT-large for routing, yes/no verification, and 1-5 rating. Non-generative single forward pass with calibrated confidence.
★34
给 WLED 灯带做的实时 shader 与节点图编辑器,把灯效从手写 JSON 预设变成连节点即时预览。玩灯带硬件的人会立刻想试,此前只有零散网页特效预设;局限是绑死 WLED 生态且刚发布。
Live shader and node graph editor for WLED Strips
★25
给 KOReader 换睡眠屏的小插件,作者直接承认灵感来自 CrossPoint,范围克制。对电纸书用户是装上就能用的十分钟项目;局限也明显,只服务 KOReader 生态,功能就是换一张待机画面。
Book Card is a new sleep screen for KOReader, heavily inspired by the idea and visual design of CrossPoint's sleep screen implementation.
★20
今天就到这里。下面是其余等级和榜单。
微信(Windows 4.x)旁挂的回复辅助:窗口截图 + 本地离线 OCR 读对方消息 → Jev 判断意图 → 3 条候选一键填入,发送永远手动
非自回归System 1决策引擎,专为结构化类型决策场景设计 DeepOpen Multilingual, non-autoregressive System 1 decision engine.
A local AI gateway for AI agents, unifying your subscriptions and API providers with smart routing and on-device privacy protection.
Android virtual camera for security testing. Injects frames into cameraserver so any app sees a video file, desktop stream, or webcam instead of the real camera.
微信聊天记录一键转发到 AI Agent 与 Obsidian 的原生 macOS 工具
Low-latency screen streaming from Apple Silicon to a 2010 iMac on OS X 10.9. Hardware H.264 over RTP/UDP, virtual display, zero-copy OpenGL rendering.
Score every test against a git diff with Jev, and emit the filter arguments vitest, node:test, Playwright, cargo test and go test already understand
Adaptive native glass surfaces for Expo and React Native. Scales blur, refraction and highlights to device capability and live frame performance.
AltServer for iOS 17+ that runs on your home server (Docker): renews and installs sideloaded iPhone apps over Wi-Fi, no computer or VPN
DolphinBench - Mapping the Pareto frontier of Agent Memory
Async-first agent harness
Native compact Nix build output with full local logs, bounded diagnostics, and a flake interface
Agent Skill: turn raw app screenshots into App Store and Google Play marketing images, UI pixel-exact
Question video transcripts, inspect the evidence, and organize clips from the terminal.
Learned 2x latent upscaling for Qwen Image 2.1 in ComfyUI.
Fan PC port of WET (Xbox 360) using ReXGlue. Does not include game files.
Self-developed dual-mode Bluetooth stack in standard C++. Cross-platform for Windows, Android, Linux and RTOS. Modular design with thread-pool scheduler. Inter-module message delivery, serial executio…
Proof of concept for CVE-2026-43786, a local privilege escalation vulnerability in macOS CoreServices that allows an application to gain root privileges.
An unofficial local visualizer for Jev (TypeSafe): a live view of every call your code makes. Not affiliated with TypeSafe AI.
Interactive 3D map of the human tech tree, from the first stone tools to 2024: technologies stacked as a tower of eras in three.js, with prerequisite and relation edges, hover cards and a guided camer…
LLM-assisted quantitative factor research with safe DSL validation, benchmark comparisons, and cost-aware Top-K backtesting
Visual Category Theory Studio — draw commutative diagrams, 2-cells, adjunctions and 3D cubes in the browser; export SVG, PNG and tikz-cd. Concept by Dmitry Vostokov, DumpAnalysis.org
A terminal coding agent that plans work as executable graphs
opencode plugin to monitor agent drift over time
MiMo-V2.6-Flash-RL on 2x NVIDIA DGX Spark (GB10): vLLM TP2 + DFlash + vision/audio, fp8 KV 300K, four fixes for the stock image, measured baselines
把一个主题做成有配音、字幕、封面的讲解分析类视频 —— 全本地、用 HTML 写画面,稳定音画同步。 面向 AI 编程智能体的确定性 HTML → MP4 渲染流水线。
Describe a YC startup in any words and the matching logos float up out of a physics pile, each with how likely it is. 6,241 companies, searchable by what they do, what their logo looks like, and what …
Procedural risograph films in single HTML files (Window Seat, Roost and more), with the Claude Code skills, docs and tools to make your own.
Jev Chat Assistant for Windows - 电脑版微信意图判断与 DeepSeek 建议回复
免登录、无广告、专注收听的第三方 B 站有声小说播放器
A lightweight, native Swift 6 bridge integrating TypeSafe AI's Jev System One decision model into Apple's Foundation Models framework.
微信消息意图识别悬浮窗(macOS):看屏 + 本地小模型判断意图和风险,再按话术生成回复候选。纯只读、不注入微信。
Adaptive reasoning effort for GPT-6 during Codex tasks, powered by Jev to reduce token usage.
📦✨ AI Builder For Creating Sites and Apps! An open-source alternative to Lovable, Replit, v0, and similar platforms.
AI Manager desktop application for AI coding tools
Defend sound research against LLM reviewer bias through meaning-preserving rewrites.
會議資訊卡 × LINE LIFF 分享 × 會議機器人:把一場會議變成一張可以在 LINE 裡轉傳的卡片(含 LIFF、部署、Claude Code 教學)
805 verified examples of Jev — TypeSafe AI's System One decision model — indexed by the decision each one makes, not the blog that mentioned it. Every cited call site is re-read by CI each week. Bilin…
Open-source, self-hosted catalogue raisonne for artists who work on-chain.
Hand-drawn art and animation, written as code. Illustrations, loops, stickers and scored films in nine styles, identical on every render.
Jev-style structured decisions on DiffusionGemma: the example server from vLLM PR 57250
Unofficial PoC: a BERT-style encoder decision engine behind a typed-decision (noul / choice / score) HTTP API. Not affiliated with TypeSafe.
Turn any LLM into a Jev-style decision model: typed decisions, real probabilities, no training.
Personal or team knowledge base, supports MCP API calls.
C-HD: Lean 4 proof package for a directed single-source shortest-path bound (snapshot 2026-09-20)
Same as jev-ultrafast but using Laya
独立开发的 Cloudflare Anycast 域动 IP 优选与自域工具
A skill for handling conflict without canned empathy, avoidance, or over-analysis.
微信(FkWeChat 插件):长按消息分析意图 / 情绪 / 回复姿态,只在本机弹提示,对方无感知
从游戏客户端反推服务端协议并复现可部署服务端 (Agent Skill)。Unity/UE/Cocos, IL2CPP/Lua/JS…
Field report: splitting one React Native app into two apps sharing a package. Metro duplicate instances, npm workspaces, native identity.
极快的ComfyUI内QwenImage2.1加速节点.
Reverse engineering analysis of PureRAT, a multi-stage RAT that executes via msbuild.exe and communicates with C2 servers at pure8s.ddnsfree.com and 52.241.248.38. Uses .NET evasion techniques (Disabl…
Open-source AI coding desktop with DeepSeek Harness and private GitHub backups of uncommitted code.
Live SEO audit for any website from one homepage URL, judged by Jev. PDF, XLSX and Markdown reports.
真实面经合集:按公司、岗位分类,包含面试流程、真题回顾与复盘
Laya vs Jev: local MLX and hosted AI decisions playing T-Rex side by side, with live metrics and replay recording
MiMo-V2.6-Flash for 2x DGX Sparks
Engineering rules for building trustworthy coding-agent workflows, inspired by Lauren Tan.
ZCode 的开源社区版:移除遥测、保留官方服务权益、提供开源文档与桌面自动化能力
对Qwen-3.5-4B这个视觉语言模型进行jev般改造,识别速度达到0.169秒/帧。切换其余Qwen模型直接可以按这个逻辑加速。
Transforme materiais de concursos em uma base de conhecimento inteligente no Google NotebookLM. Pipeline completo com automação de download e transcrição (Gran/Estratégia/QC), sincronização dinâmica v…
A Clash Royale bot with no trained policy: Jev (TypeSafe System One) makes every decision from the live game state
O ESTAGIÁRIO — IA CONTÁBIO: automação que trabalha nos sistemas que o escritório já usa.
《Jev 工程学:为 coding agent 而作》完整中文翻译 — 保留原结构与 7 张插图
interactive travel planner
Super fast Jev like server, model agnostic, working with any OpenAI compatible endpoint
Permutation Invariant Jev
Hermes Desktop computer pane stacked above Files. Fork of thomasbek3/hermes-bot-kit computer-viewer.
config files for fastfetch and starship
ngcc artifacts
Reference client daemon and verification worker for Kryvora Network nodes.
Jev shadcn-block generator
ChatGPT skills are reusable workflows made of instructions, examples and code. Where to find them, the four ways to create one, and the eligibility and scan rules.
⚡ Millisecond-level Viral Hook Telemetry & Co-pilot for Social Media (Chrome Extension + JEV System 1)
One-menu RTX 5090 setup for non-Swift Huihui Qwen3.8-27B NVFP4 with Cinference, MTP-10 and K8V4.
Lightweight MarkText companion
FDE 现场工程手册
Fast, minimal, effective
travel planning
Curated instagram username checker list: availability tools, multi-network and domain checks, next steps.
Python projects demonstrating programming fundamentals, data structures, validation, automation and application logic.
Nano GPT alternatives compared: NanoGPT vs OpenAI direct, Anthropic direct, self-hosted LiteLLM, DeepInfra and Synexa.
Streamlined desktop application
A chain for agents, on BNB Smart Chain. Genesis carries three system contracts and three neutral tools — Multicall3, a CREATE2 deployer and Wrapped BAC. Every token, pool and market beyond that is dep…
Fresh build for daily workflow
Desktop edition for power users
Seedance 2.5 doubles single-shot output to 30 seconds and raises the reference ceiling to 50 inputs. Where to run it, what it needs, and what nobody prices.
Seedance API access has numbers now: $0.1512 a second, a 124-second median render and an async task endpoint. How to read them before you wire anything up.
Fresh build for daily workflow
Modern desktop experience
Modern desktop experience
Clean interface, solid performance
Official-style desktop package
Fast, minimal, effective
Clean and fast companion tool
Codex usage is reported in three different places that do not agree. What /status, the Codex settings page and the API dashboard each actually measure.
Official-style desktop package
Streamlined desktop application
Clean interface, solid performance
Official-style desktop package
Polished and ready to use
41-module Cheat/Trainer for IGTAP: an Incremental Game That's Also a Platformer — movement tools, power economy, clone automation, progression and speedrun utilities.
答案之书jev
面向中美双栖、高收入第一代移民家庭的生活操作手册:医疗、保险、税务、法律、社会关系、父母养老、跨境生活与家庭风险管理。
Turn token activity into compute. A compute-access protocol designed for Robinhood Chain: protocol fees fund GPU capacity for token holders.
A showcase of exceptional student projects from Lords Institute of Engineering and Technology
Modern desktop experience
Everything you need in one place
Simple, focused workspace
Coreção do CP1 - Estruturas Semânticas
Automated reporting assistants are programmed software mechanisms designed to streamline and aggregate the submission of appeals or violation notices regarding Facebook profile, page, or group acc rep…
Simple, focused workspace
coming soon
Hunt Showdown 1896 Hack 2026 themed PC companion for Hunter loadouts, weapons, traits, manual aim practice, match planning and performance analytics
A Hootsuite review for people deciding in 2026: what the five-product Social OS covers, who it is priced for, and the lighter option for a small team.
Kimi API review from outside Moonshot AI: which models the site names, what the K3 page claims, where pricing lives, and when another API suits you better.
Kiro CLI puts agentic engineering in the terminal: executable specs, hooks and parallel sessions. An independent review of what V3 changed and who should skip it.
Clean interface, solid performance
Streamlined desktop application
Updated desktop experience
Cross-platform ready package
This Public Source Code Is Not Written Or Maintained By Tech VJ.
Matériel de cours M2 INALCO
Clean and fast companion tool
Reliable desktop companion
Reliable desktop companion
871 PKG games for fpkgi
Updated desktop experience
Fast, minimal, effective
Clean interface, solid performance
Reliable desktop companion
37-module Trainer/Cheat for Dressmaker — coins and shop, instant sewing and pattern cutting, order fulfilment, sketchbook design and cozy atelier tools.
rootless Linux沙箱,用namespace加gVisor给本地跑不可信依赖与第三方程序一个可丢弃的隔离home,解决装个包就整机沦陷的痛点;153分HN说明开发者共鸣强,前两次无人问津后靠gVisor支持翻身。可做AI生成代码执行底座。风险是firejail/bubblewrap与容器已占位且仅限Linux。
I created Drop because I always felt uneasy installing and running third-party programs using my main user account. A single compromised dependency means a full compromise of the system. What is even …
▲153
机器人VLA与世界模型的推理加速运行时,把扩散采样从25/50步压到2/4步,直接解决控制回路延迟这一部署瓶颈;有Jetson Thor实测与三星西门子使用背书,HN 23分但GitHub仅42星,动能刚起步。可成为机器人侧vLLM。风险是AGPL与NVIDIA Isaac/TensorRT生态挤压。
Hey HN, Guanming here, cofounder of General Instinct. We just released InstinctFlash, a high-performance serving framework for robotics models. It’s licensed under AGPL-3.0. On Jetson Thor, we see spe…
▲23
用Rete规则引擎出判决、LLM只负责引用政策解释,解决信贷风控临床这类必须可审计的决策问题,Verdict可复现是真实卖点;34分HN与MCP客户端说明有早期开发者尝鲜。商业化路径清晰(合规审计SaaS)。风险是引擎闭源、被Drools/IBM ODM及模型自带guardrail夹击。
Hi HN, I built ai·rete·rag because I kept seeing teams put an LLM in charge of decisions that need to be auditable (lending, fraud, clinical triage), then bolt on "guardrails" after the fact. It runs …
▲34
不看词只看结构判断网页是否AI生成,98%准确率且改写难以规避,解决营销团队对AI slop的检测需求;37分HN与YC背书说明有商业抓手,已内嵌Sitefire产品。路径是内容质量API。风险是检测与规避的军备竞赛,以及人类样本仅限2020至2022年。
Hey HN! We’re Vincent and Jochen from Sitefire ( https://sitefire.ai). We have been working together for years, with backgrounds in RL/optimization at Stanford and software engineering from Technical …
▲37
为Jev这类返回受限选项与概率的决策模型做可复现基准,解决非自回归模型缺乏统一评测口径的问题;53分HN说明话题有热度,公开题目与打分代码增强可信度。若被采纳可像lm-eval成为标准。风险是厂商主导、榜首即自家模型的利益冲突,公信力难维持。
Hi HN! I built JevBench because Jev kicks ass, and the world deserves to know how the serious open source and fake lookalike projects really perform in comparison. Jev-class models return bounded choi…
▲53