AI 代理接管终端、游戏测试与动画逆向：今日开源项目亮点
日报 · 2026-10-31 · GitHub 新星
覆盖日期：2026-10-31 · 北京时间
发布时间：2026/11/1 08:00:44
本期日报收录 45 个开源项目，其中应用与服务类占多数，AI 代理相关工具尤为集中。这些项目展示了 AI 从对话助手向工作流执行者的转变：从管理终端和编码代理的桌面应用，到主动编码代理原型，再到游戏自动测试、动画逆向工程和 3D 模拟生成。同时，本地优先的搜索与求职助手，以及 GPU 解锁、免 U 盘装 Linux 等硬件工具也值得关注。
今天的开源项目里，AI 代理不再只是聊天窗口里的助手，而是开始接管终端、游戏测试、动画复制甚至求职准备等具体场景。本期收录的 45 个项目中，应用与服务类占 35 个，其中不少项目都在解决同一个问题：如何让 AI 代理在真实工作流中更主动、更可靠地完成任务。除了 AI 相关工具，本期还出现了 GPU 解锁、免 U 盘装 Linux 等硬件向项目，以及本地优先的搜索与求职助手，整体覆盖了从开发者日常到普通用户的多种需求。
1. 开发者的新工作台：从终端管理到 AI 代理调度
对于每天要开多个终端、跑多个编码代理的开发者来说，Herdr 提供了一个值得关注的新思路。这是一款桌面应用，把终端和编码代理组织成“空间”（space），每个空间可以包含多个标签页和窗格，并支持通过 SSH 连接远程机器。它的核心设计是让终端在窗口关闭后继续运行，代理的状态（Working、Blocked、Done、Idle）会汇总到侧边栏，方便开发者一眼看出哪些任务需要人工介入。Herdr 还允许用 JavaScript 编写“mods”来定制界面，并提供了 herdr-app --skill 命令，让代理学习如何驱动窗口。对于同时管理本地和远程多台机器、且经常被多个代理任务打断的开发者，这款工具把“代理需要你”和“你需要的终端”整合到了一个界面里。
仓库链接：https://github.com/herdrdev/herdr-app
如果说 Herdr 面向的是熟悉终端的开发者，那么 Capri 则试图把 AI 代理带给不习惯命令行的 Windows 用户。Capri 是一个桌面 AI 代理应用，用户用自然语言描述任务，它就能调用文件、项目、代码和网页等工具自动执行。它本身不包含 AI 引擎，而是作为“身体”接入 Claude Code、Codex 或 Antigravity 等“大脑”，用户自带账号或 API 密钥。Capri 强调“无需终端”，安装后从开始菜单启动，首次运行连接 AI 助手即可。它支持文件整理、代码调试、文档汇总等常见任务，并在执行前提供审批环节。不过目前仅支持 Windows，macOS 和 Linux 版本尚未推出。
仓库链接：https://github.com/boardshogunfluctuate/Capri-AI
在更底层的层面，DeepSeek 团队开源的 dsh-superagent-prototype 展示了一种更主动的编码代理形态。它面向单个 GitHub 仓库，集成在 DeepSeek Harness 中，可以监听 GitHub 和飞书，当被授权的人提出请求时，代理会自行创建工作项、在自己的会话和目录中处理，并以拉取请求等形式回复。项目还附带一个网页面板，显示哪些事项需要人工处理。值得注意的是，项目明确标注为“研究原型，仅供参考，非官方产品”，代码由编码代理完成大部分工程，可能包含错误，工程实践也不一致。它体现的设计哲学是“尽可能少占用人的注意力，尽可能多创造价值”——代理只在被要求或团队规则允许时行动，只在必须有人行动或结果可用时才说话。对于想了解未来编码代理形态的开发者，这份原型值得研读，但直接用于生产环境前需要充分审查。
仓库链接：https://github.com/deepseek-ai/dsh-superagent-prototype
2. 让 AI 看懂游戏和动画：测试、逆向与模拟
游戏开发领域出现了几个有趣的 AI 应用。Aura Autopilot 是一个实时游戏测试系统，支持 Unreal Engine 5、Unity 6 和 Godot 4，可以在 Windows 和 macOS 上运行。它通过编辑器插件集成，能够自动探索关卡、测试功能、验证行为，并生成报告。用户可以用自然语言下达四种类型的指令：探索（如“探索 Lvl_Main 中的制作系统”）、测试（如“检查当前关卡并告诉我所有看起来不对劲的地方”）、验证（如“验证用斧头砍树时树会摇晃”）和目标（如“在生命值高于 20% 的情况下击败 Boss”）。Autopilot 会启动 PIE 窗口，读取项目源码，根据提示自动游玩，完成后关闭窗口并生成报告。它还可以通过 Claude Code 或 Codex 集成，使用自己的订阅。不过，在 Aura 中仅支持 Unreal 和 Unity，从源码安装也仅支持 Unreal Engine。
仓库链接：https://github.com/ramenvr/autopilot
动画逆向工程是另一个有趣的方向。reanimate 是一个 Claude Code 技能，可以从实时网站、屏幕录制、mp4、gif 或 Lottie 文件逆向工程动画，生成代码副本，并通过逐帧对比验证一致性。它的工作流程是：录制原始动画，在代码中重建，再以相同方式录制副本并对比，找出差异后继续修复，通常 2 到 5 轮就能收敛。对于 canvas 和 WebGL 动画，它还会追踪 GPU 调用（着色器、uniform、draw call、three.js 场景），并在虚拟时钟上保证每次捕获一致。默认情况下它匹配运动轨迹，如果要求“像素级一致”，则会进一步匹配字体、大小、颜色、间距和资源。不过，它无法直接录制需要登录的页面、YouTube 和 Loom，物理弹簧效果是通过拟合近似而非直接读取，且需要 Bun 1.1 或更高版本，Windows 未测试。
仓库链接：https://github.com/NiazMorshed2007/reanimate
teardown 则把 AI 带入了机械拆解演示。这是一个 Claude Code 技能，能把任何机器转化为可拆解的 3D 模拟器。用户用自然语言描述一台机器（比如“四冲程摩托车发动机”），Claude 会生成一个离线 HTML 页面，机器在中间运行，周围面板显示实时数据，还有一个按钮可以打开正在工作的部件，无人操作时页面会自动导览。项目内置了 V8 发动机、Cessna 337 起落架和 Strandbeest SB-10 三个示例。它需要 Node.js 20 或更高版本，生成视频需要 ffmpeg，声音需要 Python 3。对于想快速理解机械原理或制作教学演示的人来说，这是一个低门槛的工具。
仓库链接：https://github.com/elvarofficial/teardown
3. 本地优先的搜索与求职助手
Kensaku 是一款本地优先的桌面文件搜索启动器，支持按文件名、文件内容、图片中的文字（OCR）以及语义进行搜索。它声称可以在约 12 MB 内存下，用约一毫秒搜索 10 万个文件名。除了常见的 PDF、DOCX、PPTX 等文档格式，它还能读取截图中的文字（默认使用 PP-OCRv6，可选更高精度的 Indra 引擎），并支持自然语言时间查询（如“上周的截图”）。Kensaku 还提供 MCP 接口，让 Claude 或 Cursor 等 AI 代理可以搜索本地索引，默认只读。它支持 Windows、macOS 和 Linux，但构建尚未代码签名，自动更新也未启用。对于经常在大量文档和截图中找东西、又注重隐私的用户，这款工具把“搜索”从文件名扩展到了内容甚至语义。
仓库链接：https://github.com/yjid/kensaku
求职辅助是另一个本地优先的应用场景。svoi 是一款面向 Windows 的免费本地优先助手，帮助用户分析简历、查找真实职位（通过 HH.ru 开放 API）、准备和编辑回复、准备面试，以及处理就业后的工作事务。它明确声明不会自动发送回复、不绕过 CAPTCHA、不保证就业，也不将技能匹配视为邀请保证。用户需要自己确认个人资料、文本和实际发送回复。数据存储在本地加密存储中，可选的公开证书页面只发送用户选择的昵称、状态和聚合计数器。这款工具的价值在于把求职流程中的重复性工作（如根据职位要求准备面试问题）本地化处理，同时保持人对最终输出的控制。
仓库链接：https://github.com/Balanov/svoi
4. 硬件与系统工具：GPU 解锁和免 U 盘装 Linux
本期还出现了两个面向硬件玩家的工具。nv-uefi-unlocker 是一个 UEFI 预驱动 NVIDIA GPU 解锁器，在驱动加载前通过 UEFI DXE 时窗执行完整固件事务，解除 RTX 20/30/40 系显卡的 tensor 发射率钳制。项目声称在 RTX 2070 上实现了 FP32 累加 GEMM 从 36 到 60–63 TFLOPS（+67~75%），在 RTX 4060 Laptop 上 FP16 GEMM 从 29.89 到 49.33 TFLOPS（1.65×）。不过，解锁是易失的，断电或冷启动后失效，每次需要重放；RTD3 和雷电/热插拔环境会擦除解锁状态。项目明确要求用户自行承担硬件风险，并提供了详细的兼容性矩阵和操作流程。对于想挖掘旧显卡性能的玩家，这是一个高风险高回报的工具。
仓库链接：https://github.com/physicsdolphin/nv-uefi-unlocker
hop 则是一个更亲民的 Windows 应用，可以在已有 Windows 系统旁边安装 Linux，无需 U 盘。用户选择发行版、拖动调整分区大小，点击安装即可。它从官方源下载 ISO 并校验校验和，安装前会检查 PC，磁盘更改前所有检查必须通过。它支持 elementary OS、Ubuntu 及其风味版本、Linux Mint 和 Pop!_OS，但不支持 Kubuntu、Fedora 和 Arch。如果安装失败回到 Windows，重新打开 hop 会解释原因并可以重启进入 BIOS。对于想尝试 Linux 双系统但不想折腾 U 盘的新手，这款工具降低了门槛。
仓库链接：https://github.com/Arjun-Ingole/hop
5. 决策模型与网页自动化：更精细的 AI 控制
switch 是一个构建决策模型框架的工具，它让决策模型（如 Jev）通过有限选择逐步完成任务。与编码代理不同，决策模型只回答类型化问题并输出概率，不生成文本，因此无法直接调用工具。switch 在代码中生成工具调用，模型只需从有限集合中选择。它的典型应用是评估中的评判员：在需要检查大量文档的评估任务中，传统方法会把所有文本甚至图片都塞进上下文，而 switch 让决策模型逐步导航文档，只读取需要的内容。项目提供了国际象棋、维基竞速和评分评判员等示例。它需要 Node.js 22.18 或更高版本。对于研究决策模型或构建高效评估管道的开发者，这是一个值得关注的框架。
仓库链接：https://github.com/osolmaz/switch
cfsolve 是一个 Rust 库和命令行工具，用于解决 Cloudflare Turnstile 小部件和挑战页面，无需无头浏览器。它通过模拟浏览器配置文件（目前仅支持 Safari 和 Chrome/ium）来通过挑战，返回 Turnstile 令牌或 cf_clearance cookie。使用 cf_clearance 时，必须从与配置文件相同的 IP 地址和用户代理发送，否则 Cloudflare 会拒绝。构建需要 Rust、CMake、C/C++ 编译器、libclang 和 Git。这个工具适合需要自动化处理 Cloudflare 验证的开发者，但需要注意其用途可能涉及绕过网站防护，应谨慎使用。
仓库链接：https://github.com/chris124567/cfsolve
6. 结语
今天的项目呈现出几个值得注意的方向：AI 代理正在从“对话式”走向“工作流式”，它们不再只是回答问题，而是主动管理终端、测试游戏、逆向动画、搜索文件，甚至辅助求职；本地优先和用户控制成为不少项目的设计原则，无论是 Kensaku 的本地索引、svoi 的本地加密存储，还是 Capri 的审批环节，都在强调人对数据和操作的控制权。对于开发者来说，这些工具提供了将 AI 融入日常工作的新方式，但部分项目尚处于原型或未签名状态，使用前需要仔细阅读其文档和限制。
文章生成于 2026/11/1 08:01:32 · 依据本期保存的资料撰写
本期项目与来源资料
收录数按本期首次达标收录的项目统计。
- deepseek-ai/dsh-superagent-prototypeTypeScript · 2,222 Stars
代码任务自动化
Stars 统计更新于 2026/11/1 07:55:47
查看原始资料
作者描述
Towards a More Autonomous Agent System
编报时的 AI 中文解读
dsh-superagent 是一个面向单个 GitHub 仓库的主动编码代理，集成在 DeepSeek Harness 中，可监听 GitHub 和飞书，自动处理工作项并生成拉取请求。
编报时的 README 依据
dsh-superagent is a proactive coding agent for one GitHub repository, packaged as a bundle for [DeepSeek Harness](https://www.deepseek.com/en/harness/). It listens to GitHub and, optionally, Feishu. When someone allowed asks the agent for work, it opens a work item, works on it in its own agent session and directory, and answers in the original conversation, for example with a pull request. A web panel shows what needs a person.
# dsh-superagent > **Research prototype for reference only, not an official DeepSeek product.** This code grew out of our team's day-to-day development work, with much of the engineering done by coding agents. Treat it as research code, not production-ready software. It may still contain bugs, and its coding style and engineering practices may be inconsistent. Review and adapt it before using it in your own environment. >
- herdrdev/herdr-appRust · 1,199 Stars
终端管理、远程开发
Stars 统计更新于 2026/11/1 07:56:31
查看原始资料
作者描述
herdr desktop app
编报时的 AI 中文解读
Herdr 是一款桌面应用，用于组织终端和编码代理的工作空间，支持多标签页和窗格，并能远程连接其他机器管理终端。适合需要同时处理多个终端会话和代理任务的开发人员。
编报时的 README 依据
## Work on another machine Add an SSH machine from Settings → Machines or the sidebar's Add machine action. Pick a host from your SSH config or type how SSH reaches it. Spaces on that machine run their terminals there, not on your computer. Each machine keeps its spaces, tabs and splits in its core. Connected screens share that structure and the live terminals, but keep their own focus and zoom. Remote work keeps running when your local window closes or your laptop sleeps, as long as the remote machine and core stay up. ## Make it yours
Use herdr-core for running and coordinating programs; use the app to change what you see. ```sh herdr-app ls herdr-app status herdr-app commands ``` The first two inspect the open window. The command list also works without one. See the [command reference](docs/commands.md) for arguments. ## Build from source For bugs, mod API requests and code contributions, see [Contributing](CONTRIBUTING.md). Clone this repo with its submodules. Use Rust 1.95.0 and Zig 0.16.0, the versions used in [CI](.github/workflows/ci.yml). Linux needs the development packages listed in that workflow. ```sh
# Herdr Herdr is a desktop app for terminals and coding agents. Keep work in spaces, see which agents need you, and leave terminals running when you close the window. ## Install ### Linux and macOS ```sh curl -fsSL https://releases.herdr.dev/app/install.sh | sh ``` Linux needs a desktop session and a working Vulkan driver. macOS builds are available for Apple Silicon; Intel Mac users can [build from source](#build-from-source). Linux builds are x86-64 and ARM64. The installer adds Herdr to your apps and puts the command in `~/.local/bin`.
Add that folder to your PATH if the installer asks. On macOS the app lives in `~/Applications/Herdr.app`. For drag-and-drop installation on Apple Silicon, download [Herdr.dmg](https://releases.herdr.dev/app/Herdr.dmg) and drag Herdr to Applications. ### Windows Run in PowerShell on x64 Windows with curl and tar installed (Windows 10 1803 or newer): ```powershell irm https://releases.herdr.dev/app/install.ps1 | iex ``` Herdr appears in the Start menu. The installer adds `%LOCALAPPDATA%\Programs\herdr-app` to your user PATH. Open a new terminal to pick it up.
- Arjun-Ingole/hopPowerShell · 701 Stars
Linux安装、双系统安装
Stars 统计更新于 2026/11/1 07:41:20
查看原始资料
作者描述
Install Linux next to Windows without a USB stick
编报时的 AI 中文解读
hop 是一个 Windows 应用，用于在已有 Windows 系统旁边安装 Linux，无需 U 盘。用户选择发行版、调整分区大小后点击安装即可，适合希望双系统安装的普通用户。
编报时的 README 依据
<p align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="assets/brand/hop-lockup-dark.svg"> <img src="assets/brand/hop-lockup-light.svg" alt="hop" width="200"> </picture> </p> <p align="center">Install Linux next to Windows without a USB stick.</p> <p align="center"> <img src="docs/screenshot.png" alt="The hop window: install elementary OS next to Windows, with 120 GB, above a bar showing drive C: split between Windows and Linux" width="760"> </p> ## What it is
hop is a single Windows app that installs Linux alongside Windows. Pick a system, drag to choose how much space it gets, and click Install. There's no USB stick to find and nothing to install first. ## What it does - **Downloads from official sources.** It gets the ISO from the project's own server and verifies its checksum where the project publishes one. You can also use an ISO you already have. - **Makes room safely.** It checks your PC first, and nothing on the disk changes until those checks pass.
3. **Copy** the ISO's files onto a small FAT32 partition, and check each one. 4. **Add a startup entry** with the ISO's signed boot loader, for the next restart only. 5. **Restart.** The Linux installer opens and installs into the space hop made. More detail is in [How hop works](docs/ARCHITECTURE.md). ## Use it Download `hop.exe` from [Releases](../../releases) and run it. Windows asks for administrator access, and SmartScreen may warn because hop isn't code-signed yet: click **More info → Run anyway**. If something goes wrong, see [Troubleshooting](docs/TROUBLESHOOTING.md). ## Build it
**Needs:** Windows 10 (1803 or newer) or 11, on a UEFI PC with a GPT disk, which is nearly every PC sold with them. ## How it works A USB installer is just a FAT32 drive with the ISO's files on it. hop makes the same thing on your internal disk: ``` Before [ EFI ][ C: Windows ][ Recovery ] After [ EFI ][ C: Windows ][ Installer ][ Linux ][ Recovery ] ``` 1. **Download and check** the ISO, and confirm it's a system hop can install. 2. **Make room** after C:, using free space first and shrinking C: only if needed.
- **Tells you what to click.** It lists the exact choices to make in the Linux installer. - **Gets you unstuck.** If the PC goes back to Windows, reopening hop explains why and can restart into the BIOS. - **Cleans up after itself.** Undo gives the space back to Windows, and it can also remove another Linux already on the drive. **Systems:** elementary OS, Ubuntu and its flavours (Budgie, Xubuntu, Lubuntu, Cinnamon, Unity, Edubuntu, Studio), Linux Mint and Pop!_OS. Other Ubuntu-based ISOs should work too. Kubuntu, Fedora and Arch aren't supported.
- physicsdolphin/nv-uefi-unlockerRust · 620 Stars
GPU解锁
Stars 统计更新于 2026/11/1 07:46:59
查看原始资料
作者描述
UEFI pre-driver NVIDIA GPU unlocker for Windows Linux Hackintosh and all EFI-bootable system (Rust): SEC2 scrubber -> FWSEC/WPR2 -> signed Booter -> V67 four-write payload, read-back verified. Verified on most RTX 20 30 40 Series.
编报时的 AI 中文解读
这是为 Windows、Linux、Hackintosh 等所有 EFI 可引导系统设计的 UEFI 预驱动 NVIDIA GPU 解锁工具，在 NVIDIA 驱动加载前通过 UEFI DXE 时窗执行完整固件事务，解除 20/30/40 系显卡的 tensor 发射率钳制。主要面向动手能力强的硬件玩家和超频爱好者，需要自行承担硬件风险。
编报时的 README 依据
| 上次离开方式 | 本次启动后的 dGPU | 本工具 | | --- | --- | --- | | 混合模式 · 纯 iGPU 关机 | 冷：无 GOP，BAR0 全 `0xffffffff` | ❌ claim 前置即报 "dGPU powered off (D3cold)" | | 混合模式 · 混合关机 → 重启 | 热：有 GOP，寄存器可读 | ✅ 全链跑通（四写 + 选择器穿 FLR，即 evidence 里的成功日志） | | 混合模式 · 休眠/唤醒之后 · D3 门控态 | 停在 D3、域门控（VGA_WORKSPACE 读回 `0xbadf1301`） | ❌ claim 前置即报 | | 混合模式 · 休眠/唤醒之后 · 事务可跑 | 寄存器可读、全链能绿，但卡是热的 | ⚠️ 事务绿 ≠ 解锁有效：出口后可能被作废，先看 `HANDOFF-OUT` 并做 Windows 侧读回（`docs/warm-resume-entry.md`） |
| 混合模式 · 休眠/唤醒之后 · **TU106 实测** | 寄存器可读、全链绿 | ✅ **写入解锁成功且 Windows 验证通过**（2026-10-27，2070 台式实测：休眠恢复 → 进 U 盘 → 全链 → 解锁存活进 Windows）。上两行的"作废"警告在 TU106 + GSP 模式组合下未复现 | | 独显直连 · 休眠或重启 | 始终是 panel owner、始终有电 | ✅ 均可用 | 操作建议：**从 Windows 选"重启"**（或冷启后直接从启动菜单进 U 盘）再运行；不要用 "关机再开"的 iGPU-only 路径，也不要在休眠唤醒后直接进 U 盘。已实现 PM 预检：入口 非 D0 会打 `PM: WARM ENTRY` 并把卡拉回 D0（读回确认），要更严格就在 U 盘根目录放 空文件 `\tx-require-d0.txt`——此时 warm 入口直接 abort、零写入零复位。 **原因分析（不是模式好坏，而是固件为谁无条件跑 POST）** 固件只在"自己要用来显示的那块 GPU"上无条件执行 VBIOS POST/devinit： - 独显直连时 dGPU 是 panel owner。冷启、热重启、S3/S4 唤醒后的任何路径，固件都必须
四写解锁双端验证：固件侧 marker `0x67b00001` + 四 PLM 读回，Windows 侧 FP16 GEMM 29.89 → 49.33 TFLOPS（1.65×）。完整报告见 [docs/ada28e0-uefi-report.md](docs/ada28e0-uefi-report.md)。 - **RTX 16 系（Turing）理论兼容**，欢迎大家实机测试。 三代机制差异见 [兼容性矩阵](#兼容性矩阵)（Turing 的 FEAT 块在 0x4096xx、 FLR 会重锁 PLM、Windows 驱动需 GSP 模式——均与 Ada 不同）。 ## 使用（RTX 40 系笔记本/台式，AD10x 家族） ### U 盘根布局（FAT32；`make-usbroot.ps1` 自动产出，手摆按下表） | 文件 | 作用 | 必需 | | --- | --- | --- | | `\EFI\BOOT\BOOTX64.EFI` | 本工具（先备份该路径原有文件） | ✅ | | `booter_load_ad102.bin` | AD102-family Booter bundle（BTL1） | ✅ |
不放 `v67-*.bin` 即为**回放/诊断模式**：走完全链但不执行解锁载荷。它仍会断开驱动、启动签名固件，并可能复位 GPU，属于写事务；只读采样应使用 `cargo efi` 构建。日志中 `family 'xxx' transaction track not enabled` 表示该芯片族尚未开放事务（20/30/40 三族均已接入；缺资产则跳过）。 **多卡**：机内多张受支持的卡时，缺省按固件枚举顺序**逐张处理**——一张 已解锁（EALREADY）或不受支持会自动让位给下一张，单张失败只对那张卡做 emergency FLR、不阻断后续；控制台在另一张卡上时全程有屏（GOP 只断目标 卡的）。`\tx-select-bdf.txt` 钉死单卡：内容与哪张卡对不上则拒绝猜测、 本轮零事务（钉的卡不在了也绝不默默跑别的卡）。 ### 操作流程 1. BIOS：关 Secure Boot。显示模式/进入方式按下面的实测表选——**推荐独显直连**， 混合模式只在"重启、且 dGPU 有电"这一种进入方式下可用（见原因分析）。 2. U 盘启动 → 版本屏 → **空格/回车确认**（其他键均中止；30 秒无操作自动回 Windows）。若 U 盘根
UEFI DXE 时窗执行完整固件事务：SEC2 scrubber → FWSEC-FRTS（WPR2）→ 签名 Booter 启动 GSP-RM →（门控后）V67 四写 payload → 读回验证 → FLR → 链式交回 Windows Boot Manager。 **当前状态：40 系（AD10x）+ 30 系（GA10x）+ 20 系（TU10x）三轨闭环—— v2.1.0 起 30 系全系兼容。** 支持矩阵： - **RTX 20 系（Turing）✅ 全系兼容 实测支持（v0.2.0）**——RTX 2070（TU106, `10de:1f07`） 全链解锁：**FP32 累加 GEMM 36 → 60–63 TFLOPS（+67~75%）**；2060 Super （`10de:1f06`）社区实测同链通过（[issue #12](https://github.com/physicsdolphin/nv-uefi-unlocker/issues/12)）。 **Turing 用户 必须先运行 [GSP 固件模式脚本](#turing20-系专属gsp-固件模式前置必读)并重启**， 详见下方 Turing 专属节。
存在 `i_know_what_i_am_doing.txt`（武装模式），确认环节变为 **5 秒倒计时后 自动运行**，倒计时内按任意键终止——文件本身就是明确的运行意图声明，供 自动化/无人值守场景使用；缺省（无此文件）永远是"必须有人按键确认"。 3. 之后全自动：事务完成或失败都会自动链回 bootmgfw 进 Windows，全程日志 落 U 盘 `\loader.log`。引导器查找顺序固定为 **bootmgfw → 各发行版 shim/grub/systemd-boot/Limine → U 盘 `\EFI\BOOT\BOOTX64.EFI`**；只有走到最后 这一档（即本工具自己所在的通用路径）才给 5 秒窗口，按任意键回固件 shell（此时解锁已完成、日志已落盘，可从 shell 手动启动任何系统）。 双系统用户可用 `\chain-target.txt` 把自定义 EFI 路径插到最前（见上表）。 4. 每次冷启动重放一次（解锁易失，无持久化、无刷写、断电即清）。 V67 payload 与全部资产再生成：`cargo run -p assets -- generate-ad102` / `release/make-usbroot.ps1`（从钉 SHA-256 的 610.43.03 公开输入本地再生，
重新初始化它，dGPU 一直有电、有 GOP、寄存器域可读——所以这里休眠和重启都可用。 - 混合模式下 iGPU 能承担显示，固件对 dGPU 只是尽力而为：冷启与混合重启路径会 POST 它（因此"混合关机→重启"可用，这也正是那份 4060 成功日志的状态）；但 - 纯 iGPU 关机会让 EC 在 POST 之前就断掉 dGPU 供电（D3cold），BAR0 全 1； - 休眠走的是 resume 路径而不是 POST：固件不重跑 devinit，EC 把 dGPU 停在 D3， PRI 域门控——VGA_WORKSPACE 读回 `0xbadf1301` 就是它在寄存器层的投影。 - 本工具在驱动加载前工作，没有 ACPI 解释器，无法像 Windows 驱动那样跑 `_PS0`/PR3 给卡上电，因此这三种失败态只能靠换进入方式规避，不能靠代码绕过。 ## Linux 用户注意：runtime D3（RTD3）会在交接后擦掉解锁（issue #9） 解锁是易失覆盖（设计上无持久化、断电即清），而 **D3cold 对相关电源域等价于 断电**。Linux 下若 NVIDIA 驱动启用了 runtime D3（RTD3），GPU 一空闲就会被
2026-10-28 用户反馈：雷电拓展坞下 eGPU 在 EFI 阶段全链解锁成功（读回全绿）， 进入 Windows 后解锁丢失。机理上这是环境问题而不是事务问题： - EFI 阶段能解锁，是因为固件在 POST 时已建好 PCIe 隧道，eGPU 对 DXE 可见； - 进入 Windows 后，TB 连接管理器（固件 ICM / Windows TB 栈的审批与隧道重建） 可能**拆除并重建隧道**，对 eGPU 等价于一次 hot reset / PERST#；坞内供电管理 （TB 控制器 D3cold、S0ix、enclosure 供电策略）还可能直接断电。解锁是易失 覆盖，任何 conventional reset（hot reset/SBR、PERST#）或失电都会把它擦掉 ——与 issue #9 的 RTD3/D3cold 同属一族。 参考系：cmp170hx 实测 Thunderbolt 3 eGPU 连 kmod 解锁本体都失败 （`Booter Load 0x15`）；Oculink 可用（本质是延长线）。**受支持环境以原生 PCIe 插槽 / 独显直连为准**；雷电/热插拔根口用户请把"解锁后进 OS 无提升" 视为预期行为，不必提 issue。
- boardshogunfluctuate/Capri-AI未标注语言 · 516 Stars
AI辅助编程、文件管理、任务自动化
Stars 统计更新于 2026/11/1 07:41:22
查看原始资料
作者描述
⚡ Tell your PC what to do. Capri is a cyberpunk AI agent that works with your files, projects and code through Claude Code, Codex or Antigravity. Windows installer, no terminal. Fan-made, not affiliated with CD PROJEKT RED.
编报时的 AI 中文解读
Capri 是一款面向 Windows 桌面的 AI 代理应用，用户通过自然语言描述任务，它可调用文件、项目、代码和网页等工具自动执行，并支持接入 Claude Code、Codex 或 Antigravity 等 AI 引擎。适合希望以图形界面方式使用 AI 代理、不愿依赖终端的用户。
编报时的 README 依据
</details> ### 3. Connect your AI On first launch, Capri asks which assistant to use. Sign in to your account or paste your API key for **Claude Code**, **Codex**, **Antigravity** or another supported assistant. ## What you can ask | You say | Capri does | |---|---| | *"Clean up my Downloads folder and group everything by type."* | Scans the folder, proposes a plan and moves files after your approval. | | *"Find the bug in this project and fix it."* | Reads the code, locates the issue, patches it and runs the checks. |
| *"Summarize every PDF in this folder into one document."* | Reads each file and writes a single combined summary. | | *"Set up a new web app and start it."* | Scaffolds the project, installs dependencies and launches it. | | *"Every Monday, back up my Documents folder."* | Creates a recurring automation for the task. | <img src="assets/divider.svg" alt="" width="100%"/> ## How it works ```mermaid flowchart LR U([You]) -->|plain language| C[Capri Desktop] C --> R{Engine Router} R --> E1[Claude Code] R --> E2[Codex] R --> E3[Antigravity]
<td width="33%" valign="top"> <b>🖥️ Desktop-native</b><br/><br/> A real Windows app with a clean interface, built for people who don't want to live in a terminal. </td> <td width="33%" valign="top"> <b>📂 Files and projects</b><br/><br/> Find, read, organize and edit files and whole project folders with one sentence. </td> </tr> <tr> <td width="33%" valign="top"> <b>⚙️ Full agent toolkit</b><br/><br/> Capri plans a task, uses tools, checks the result and reports back. Not just a chat window. </td>
Capri is open source. You use your own AI assistant account or API key, so any cost from the AI provider is billed to you by that provider. </details> <details> <summary><b>Does Capri work on macOS or Linux?</b></summary> <br/> Not yet. Capri is currently available for Windows as an installer (`.exe`). macOS and Linux versions are planned, so star the repo to follow the progress. </details> <details> <summary><b>Do I need to know how to code?</b></summary> <br/> No. Download the app, connect your AI and describe what you need in plain language. </details> <details>
| 🪟 **Installer** | [`Capri-Setup.exe`](https://github.com/boardshogunfluctuate/capri-ai/releases/latest) | Regular installation with a Start menu shortcut | ### 2. Install and launch Run `Capri-Setup.exe`, follow the wizard, then open **Capri** from the Start menu. <details> <summary><b>Windows shows a SmartScreen warning?</b></summary> <br/> New apps that are not yet widely downloaded can trigger a SmartScreen prompt. If you downloaded Capri from this repository's [Releases](https://github.com/boardshogunfluctuate/capri-ai/releases/latest) page, click **More info → Run anyway**.
| **Engines** | Claude Code · Codex · Antigravity · custom adapters | | **Platform** | Windows 10 or newer | | **Format** | Installer (`.exe`) | | **Interface** | Native desktop app, cyberpunk theme | | **Accounts** | Bring your own AI assistant account or API key | | **License** | MIT | --> <img src="assets/divider.svg" alt="" width="100%"/> ## Features <table> <tr> <td width="33%" valign="top"> <b>🧠 Bring any brain</b><br/><br/> Connect Claude Code, Codex, Antigravity or another assistant and switch between them at any time. </td>
- ramenvr/autopilotPython · 514 Stars
游戏自动测试、游戏探索、游戏验证
Stars 统计更新于 2026/11/1 07:41:22
查看原始资料
作者描述
Play any game
编报时的 AI 中文解读
Aura Autopilot 是一个实时游戏测试系统，能够自动探索、测试和验证游戏，支持 Unreal Engine 5、Unity 6 和 Godot 4。它通过编辑器插件或命令行工具集成，适用于游戏开发者和测试人员。
编报时的 README 依据
cd autopilot powershell -ExecutionPolicy Bypass -File tools\build_unreal.ps1 ` -EngineDir "C:\path\to\UnrealEngine" ` -PluginsDir "C:\path\to\YourProject\Plugins" ``` Once that has completed, follow the Autopilot setup flow as shown in Step 3 of [Installing pre-built packages](./README.md#installing-pre-built-packages). ## Usage The main way to use Autopilot is through the **Autopilot plugin** in your editor. This works in Unreal Engine 5, Unity 6, and Godot 4 on both Windows and MacOS. This path uses your Aura plan, or your BYOK API keys. Alternative ways to use Autopilot include:
- **In Aura**: You can also use Autopilot in Aura. Only Unreal and Unity are supported through this path. This path uses your Aura plan. - **In Claude Code / Codex**: You can use Autopilot through Claude Code and Codex via Autopilot's skills and MCP server. To set this up, run `autopilot connect claude` and/or `autopilot connect codex` in your terminal, then prompt Claude Code / Codex to "Use Autopilot to ...". This path uses your own Claude/ChatGPT subscription. An Autopilot prompt is one of 4 types: - **Exploration**, e.g. "Explore the crafting system in Lvl_Main."
# Aura Autopilot > By the developers of Aura, [www.tryaura.dev](https://www.tryaura.dev/)  Autopilot is a continually learning real-time general playtesting system. It supports Unreal Engine 5, Unity 6, and Godot 4 on both Windows and MacOS. ## Installation You can find the installation instructions in the [Autopilot docs](https://www.tryaura.dev/documentation/autopilot/) ### Installing pre-built packages 1. Download the latest Autopilot build for your game engine and OS from https://www.tryaura.dev/autopilot/.
1. We recommend using Autopilot with your Aura account. Click `Create Account` if you don't yet have an Aura account, or `Sign in` if you are a returning user. 2. Autopilot runs its setup, and finally redirects you to the signup/signin page. Complete your signin/signup process. 3. Upon successful authentication, you should see the "Agent service ready" page. Click Continue. 4. You have the option to enable/disable sharing your runs to improve Autopilot. As the developers of Autopilot, we would of course appreciate your enabling this!
5. (Optional) Run the 10-minute exploration run to populate your Autopilot project memory. It is safe to skip this and jump straight to sending your own prompt. 6. Prompt Autopilot! See the types of prompts you can try in [Usage](./README.md#usage). ### Installing from source You can install Autopilot from source. This is supported for Unreal Engine only. This method allows you to build Autopilot against an Unreal source build, and places the Autopilot plugin into your Unreal project's Plugins folder automatically. ```powershell git clone https://github.com/ramenvr/autopilot.git
- NiazMorshed2007/reanimateJavaScript · 454 Stars
动画逆向工程、动画复制、动画验证
Stars 统计更新于 2026/11/1 07:41:22
查看原始资料
作者描述
A Claude Code skill that reverse-engineers any animation frame by frame, then proves the copy moves the same.
编报时的 AI 中文解读
这是一个Claude Code技能，用于逆向工程网站或视频中的动画，生成代码副本，并通过逐帧对比验证动画是否一致。适合希望复制或学习现有动画效果的开发者。
编报时的 README 依据
The [skills CLI](https://skills.sh) asks which agents to set up; add `-a claude-code` to skip the question. Then, in Claude Code, say what you want, where it is and where it should go: ```text /reanimate the hero entrance on https://example.com, into components/Hero.tsx /reanimate this loader, 0:03 to 0:05 of ~/Downloads/loader.mov /reanimate the card hover in this gif, match it exactly /reanimate the logo marquee from https://example.com, with our logos and pause on hover ```
Needs [Bun](https://bun.sh) 1.1 or later, on macOS or Linux (Windows is untested). The first run installs a static ffmpeg into the skill's folder and Playwright's Chromium into Playwright's usual cache, about 250 MB together. No API keys. Other agents that read `SKILL.md` can install it too, but it has only been tested with Claude Code. ## How it works
On live stripe.com this recovers the issuing card's four-pass composite, its shaders and constants such as `u_twistFrequencyX = 0.059`. It also records the card's tilt as a curve: −0.174 → 0.262 rad between 33 and 1800 ms on the pointer path. `gl.ts compare` then checks the clone against it channel by channel; the tilt matched within 33 ms and 0.5 % of its range. See [`references/webgl.md`](skills/reanimate/references/webgl.md). ## Limits - Pages behind a login, YouTube and Loom can't be recorded directly; hand it a screen recording.
- Physics springs are fitted from their overshoot and settle time, not read from the source. - Workers are seeded and traced, but their own clocks run in real time. Canvases inside iframes need the iframe's own URL. - When a clone can't reproduce the reference's random generator, particles are compared with a blurred mismatch (density, colour, motion) instead of dot by dot. - `--gpu hardware` uses the real GPU (macOS: Metal; Linux: Vulkan, untested). The default, SwiftShader, is slower but gives the same pixels on every machine.
- chris124567/cfsolveRust · 454 Stars
网页自动化
Stars 统计更新于 2026/11/1 07:46:59
查看原始资料
作者描述
Cloudflare turnstile and challenge page solver in Rust (full reverse, request based)
编报时的 AI 中文解读
cfsolve 是一个 Rust 库和命令行工具，用于解决 Cloudflare Turnstile 小部件和挑战页面，无需无头浏览器。它返回 Turnstile 令牌或 cf_clearance cookie，适用于需要自动化处理 Cloudflare 验证的开发者。
编报时的 README 依据
You need the Rust version pinned in `rust-toolchain.toml`, plus CMake, a C/C++ compiler, libclang, and Git (to patch and build BoringSSL). ```sh cargo build --release --bin challenge ``` ## Command line tool ```sh ./target/release/challenge --page 'https://example.com/login' /tmp/solve \ --solve --fingerprint profiles/safari-26-ios.json ``` On success it prints the token (or `cf_clearance=...`) and sends progress to stderr. The directory argument holds all requests and responses for debugging, so it must be empty or not exist yet.
| `--proxy <url>` | Send every request through an `http://`, `https://`, `socks5://`, or `socks5h://` proxy. Put credentials in the URL, as in `socks5h://user:password@host:1080`. | | `-h`, `--help` | Show help. | ## Library Add cfsolve as a path dependency and copy its `[patch]` section, since Cargo only reads patches from the top-level `Cargo.toml`: ```toml [dependencies] cfsolve = { path = "/path/to/cfsolve", default-features = false } tokio = { version = "1", features = ["macros", "rt-multi-thread"] } [patch.crates-io] btls-sys = { path = "/path/to/cfsolve/vendor/btls-sys" }
included: `profiles/safari-26-ios.json` (Safari on iPhone). Each profile references recorded responses to the challenge's browser checks, stored as JSON in a directory. Keep each profile with its directory when you copy. Producing these files requires some manual (or AI) effort though it can be greatly aided by solving challenges in a real web browser with the proxy set to the address of a running `capture-proxy`, which records the decoded payloads sent to Cloudflare. Currently only Safari and Chrome(ium) profile types are supported. ## Building
as the profile, or Cloudflare will reject it. ### Turnstile with a sitekey or your own HTML `PageSource::Navigate` fetches the page. With only a URL and sitekey, use `PageSource::Blank` and specify the widget: ```rust use cfsolve::{PageSource, Turnstile, Widget}; let turnstile = Turnstile { url: "https://example.com/login".parse()?, page: PageSource::Blank, widget: Some(Widget { sitekey: "SITEKEY".into(), action: Some("login".into()), cdata: None, }), }; let token = browser.solve_turnstile(turnstile, &options).await?; ```
- elvarofficial/teardownJavaScript · 387 Stars
3D模拟生成、机械拆解演示、视频生成
Stars 统计更新于 2026/11/1 07:41:22
查看原始资料
作者描述
A Claude Code skill that turns any machine into a 3D simulator you can take apart
编报时的 AI 中文解读
这是一个Claude Code技能，能将任何机器转化为可拆解的3D模拟器，用户可用自然语言描述机器，生成离线HTML页面，包含交互式模拟、实时数据和自动导览功能。
编报时的 README 依据
npx skills add elvarofficial/teardown -g -a claude-code ``` Then, once, in the installed folder: ``` npm install && npx playwright install chromium ``` You need Node.js 20 or newer. For the video you also need `ffmpeg`, and for the sound Python 3. ## Use Ask Claude Code in plain words: > Build me a simulator of a 4-stroke motorcycle engine I can take apart. > How does a mechanical watch keep time? Show me as a simulator. > Make a teardown of a hydraulic excavator arm and record a 20-second video of it. Or open one of the 3 examples first: ```
node sims/build.mjs v8-live # teardown-out/dist/v8-live.html node sims/build.mjs cessna-live node sims/build.mjs strandbeest-live ``` ## What is inside | | | |---|---| | `SKILL.md` | What Claude reads: how a machine is built, the rules that make it look right, the review loop. | | `sims/kit/` | Runtime, stage, shader warm-up. | | `sims/v8-*`, `sims/cessna-*`, `sims/strandbeest-*` | 3 complete machines. `-shared` is the machine, `-live` the page you drive, `-film` the recorded tour. | | `sims/build.mjs` | Bundles a machine into one HTML file, three.js and fonts inlined. |
- osolmaz/switchTypeScript · 383 Stars
决策模型框架、评估判断
Stars 统计更新于 2026/11/1 07:46:59
查看原始资料
作者描述
A framework for building decision-model harnesses for any kind of task. Your program offers the model a short list of next steps, and the model picks one at a time until it can answer.
编报时的 AI 中文解读
switch 是一个用于构建决策模型（如 Jev）的框架，通过将任务状态转化为类型化问题，让模型逐步选择下一步，适用于评估、棋类游戏等场景。
编报时的 README 依据
text of the PDF, Word and Markdown files and puts all of it into the context. When the visual formatting of the output is a criterion, every page has to go in as an image too. With 200 pages and a question that needs only a few of them, most of that context is wasted. With switch, the decision model navigates the documents and takes in only what it needs. The [rubric judge](examples/rubric-judge/README.md) works this way. It reads parts of files, looks at page images and searches for terms, one step at a time, until it can grade each rubric item.
- [Rubric judge](examples/rubric-judge/README.md) grades a saved answer to a task against the task's rubric. - [Chess](examples/chess/README.md) plays chess against random moves, Stockfish, or itself. - [WikiRace](examples/wikirace/README.md) races from one Wikipedia article to another by following links. ## Write a domain A domain gives `initial`, `questions` and `transition`, and optionally `view`. This one routes a support ticket. The model reads logs one at a time until it picks stop, and then a last step asks which team gets the ticket. ```ts import { type Choice,
A run also ends when a request fails, or at a step limit that a domain may set as a safety net. [docs/MATH.md](docs/MATH.md) writes the loop down in symbols. ## Install switch needs Node.js 22.18 or newer. Add it to a project from npm: ```bash npm install @osolmaz/switch ``` | Import | What it gives | | ------------------------- | -------------------------------------------------------------------- | | `@osolmaz/switch` | The loop, the domain types, the pick helper, traces and replay |
| `@osolmaz/switch/models` | `piAiDecider`, which asks a classifier model that pi-ai lists | | `@osolmaz/switch/cli` | `commandLine`, which gives a harness the `run` and `replay` commands | | `@osolmaz/switch/testing` | A fake decider for tests that make no model calls | The examples and the [viewer](viewer/README.md) live in this repository and are not on npm. To run them, clone it and install its workspace with [pnpm](https://pnpm.io): ```bash git clone https://github.com/osolmaz/switch.git cd switch pnpm install ``` ## Examples
- Balanov/svoiJavaScript · 377 Stars
求职辅助、简历分析、面试准备
Stars 统计更新于 2026/11/1 07:41:22
查看原始资料
作者描述
Free local-first Windows companion for job search and everyday work, with editable responses.
编报时的 AI 中文解读
这是一款面向Windows用户的免费本地优先助手，用于求职和日常工作，可协助分析简历、查找真实职位、准备和编辑回复、准备面试，并处理工作事务。用户需自行确认个人资料、文本和实际发送回复。
编报时的 README 依据
# Свой **Бесплатный локальный помощник для поиска работы и повседневных рабочих задач на Windows.** «Свой» помогает разобрать резюме, найти реальные вакансии, подготовить и отредактировать отклик, подготовиться к интервью, а после трудоустройства — разбирать рабочие задачи. Пользователь подтверждает профиль, тексты и фактическую отправку отклика сам. [Открыть сайт](https://svoi.balanov.pro/) · [Скачать для Windows](https://svoi.balanov.pro/downloads/Svoi-Setup-Windows-1.0.0-x64.exe)
[Исходный код на GitHub](https://github.com/Balanov/svoi) · [GitHub Release 1.0.0](https://github.com/Balanov/svoi/releases/tag/1.0.0)  ## Что реализовано - Лёгкий онбординг: сайты поиска, резюме и прошлые отклики, подтверждение распознанного профиля. - Импорт TXT, Markdown, CSV, JSON, PDF и DOCX в desktop-версии. - Реальные вакансии HH.ru через открытый API и публичные страницы: зарплата, навыки и полное описание выбранной позиции. - Добавление вакансии по ссылке с других сайтов — без выдумывания поддержки неизвестного источника.
- Черновики откликов и ответов работодателю с редактированием; статус отправки отмечает человек. - Подготовка к интервью по требованиям конкретной вакансии. - Рабочие задачи, письма и диалог после трудоустройства. - Экспорт резюме в Markdown. - Опциональный свободный диалог с локальной Qwen3-4B или существующей Ollama. Основные сценарии доступны и без загрузки модели. - Добровольная публичная страница и QR с псевдонимом, статусом и счётчиками действий. Её можно обновить, отключить или полностью удалить.  ## Что приложение не делает
- Не отправляет отклики автоматически и не обходит CAPTCHA или ограничения сайтов. - Не обещает трудоустройство и не выдаёт совпадение навыков за гарантию приглашения. - Не публикует резюме, переписку, названия работодателей или документы в сертификате. - Не выдаёт дневник использования за удостоверение квалификации или подтверждение результата найма. ## Локальные данные и сертификат Профиль, документы, отклики и переписка хранятся на компьютере в зашифрованном хранилище текущего пользователя Windows. Поиск HH и загрузка модели требуют интернет.
- yjid/kensakuRust · 353 Stars
本地文件搜索、OCR文字识别、语义搜索
Stars 统计更新于 2026/11/1 07:46:59
查看原始资料
作者描述
Search inside documents, code, and even screenshots - instantly and locally. A free, open-source file launcher with OCR, semantic search, and MCP agent access.
编报时的 AI 中文解读
Kensaku 是一款本地优先的桌面文件搜索启动器，支持按文件名、文件内容、图片中的文字（OCR）以及语义进行快速搜索，适用于需要快速查找本地文档、代码和截图的个人用户。
编报时的 README 依据
<img src="assets/hero.png" width="100%" alt="Kensaku — find anything on your computer" /> </p> --- ## Find it the way you remember it You don't remember the filename. You remember what was in it. - The PDF about database indexing, saved as `notes-final-2.pdf`. - The log line from the failed deploy, somewhere in a folder of logs. - The screenshot with a CUDA error, which has no useful name at all. Type what you remember. Kensaku searches names, text inside files, and meaning, and shows the match with the words highlighted. Open it with Enter. ## What Kensaku does that the others don't
higher-quality Indra engine (deskew, multi-script recognition, ~27 MB pack downloaded on demand) and classic Tesseract are both selectable in Setup/Settings. **Precise** - Filters and sort: Best match, Newest, Oldest, Largest, Name. - Time travel: "last Tuesday", "3 days ago", "last week", "today", "yesterday". - A query language when you want it (below). **A real launcher** - Starts hidden in the tray and summons with a keystroke. Recents are ranked by what you actually opened, and they survive restarts.
- **macOS:** open the app, then go to System Settings → Privacy & Security and click **Open Anyway**. [docs/INSTALL.md](docs/INSTALL.md) has the full steps. **First run:** choose the folders you want indexed, pick exclusions, and optionally download the OCR or model packs. Search works while the first scan runs. **Command line:** `kensaku-cli` ships with each release. ```bash kensaku-cli search "docker error" kensaku-cli search "invoice" --ext pdf --since 30d kensaku-cli search "resume" --json kensaku-cli state ``` ### Build from source ```bash git clone https://github.com/yjid/kensaku
| Platform | Download | |---|---| | Windows | `Kensaku_<version>_x64-setup.exe` or `.msi` | | macOS (Apple silicon and Intel) | `Kensaku_<version>_universal.dmg` | | Linux | `.AppImage`, `.deb`, or `.rpm` | Every release also ships `kensaku-cli` for the command line and a `SHA256SUMS` file so you can verify what you downloaded. Builds are not yet code-signed, so macOS and Windows show a warning the first time. That is expected for an independent app, not a sign of a problem. - **Windows:** on the SmartScreen prompt, click **More info** → **Run anyway**.
cd kensaku bun install bun run tauri build ``` You need Rust (stable), Bun 1.1 or later, and your platform's webview libraries. See [docs/INSTALL.md](docs/INSTALL.md). A `Makefile` covers the common tasks: `make dev`, `make build`, `make test`, and `make reset`. ## Connect AI agents (MCP) Point an MCP client at the CLI: ```json { "mcpServers": { "kensaku": { "command": "kensaku-cli", "args": ["mcp"] } } } ``` For clients that connect over HTTP, run `kensaku-cli serve`. It listens on loopback only and requires a bearer token stored in your config directory.
- Agent access over MCP and loopback HTTP, with one-click setup for Claude and Cursor - `.gitignore` and `.cursorignore` respected by default **Known limitations** - Builds are not yet code-signed, and the auto-updater is not wired up yet. - OCR: PP-OCRv6 by default (bundled runtime, ~6 MB model pack downloaded on first run); optional Indra engine with deskew and multi-script recognition (~27 MB pack, downloaded on demand). Intel Macs fall back to Tesseract. JPEG 2000, fax and JBIG2 scans are not decoded yet. - Thumbnails cover image files only. Documents and code show type icons.
- N1nEmAn/sapper-v2Python · 336 Stars
自主渗透测试
查看原始资料
作者描述
Autonomous pentest legion — AI commander + worker swarm + layered fact-intent graph. 3rd Tencent Intelligent Attack-Defense online champion (634 teams · 8h unattended · 1B tokens beat 4.5B). The Nüwa Test: give AIs Create/Language/History, let a civilization grow.
编报时的 AI 中文解读
sapper-v2 是一个自主渗透测试系统，由AI指挥官、工人群和分层事实意图图组成，在腾讯云智能攻防挑战赛中获得冠军，适用于授权安全测试场景。
- SAOS-3624/dsh-video-storyboardJavaScript · 323 Stars
视频分镜、视频理解
查看原始资料
作者描述
Scene-aware storyboard for DeepSeek Harness: 9-16 numbered cells with timestamps + transcript + an audio envelope, packed into ONE image (384 tokens). Two tools: video_storyboard and video_zoom.
编报时的 AI 中文解读
为 DeepSeek Harness 提供场景感知视频分镜工具，将视频压缩为一张宫格图与带时间码的文本清单，供多模态模型理解视频内容。
- Balanov/baiagent-directorJavaScript · 305 Stars
商业研究、试点规划
查看原始资料
作者描述
Free local-first AI business advisor with evidence-linked research and verifiable pilot plans.
编报时的 AI 中文解读
免费本地优先的AI商业顾问，输入公司网站即可研究公开信息，区分事实与假设，并生成可验证的试点计划。适用于需要快速了解公司、市场或制定下一步行动的企业用户。
- gothamor/gta-vcs-recompC++ · 305 Stars
游戏移植、游戏增强
查看原始资料
作者描述
A native Windows and Linux port of Grand Theft Auto: Vice City Stories (PS2, USA, SLUS-21590), made by recompiling the game's PS2 executable into C++
编报时的 AI 中文解读
这是一个将PS2游戏《侠盗猎车手：罪恶都市故事》静态重编译为原生Windows和Linux移植版的项目。用户需要自备游戏ISO文件，通过构建工具在本地编译生成可运行的游戏。适合希望在现代PC上以更高帧率、宽屏和增强画质游玩该游戏的玩家。
- joeblau/subvoxSwift · 298 Stars
语音转文本
查看原始资料
编报时的 AI 中文解读
这是一个macOS原型应用，用于测试AirPods Pro头动数据能否将无声的口语转化为文字。它不使用语言模型，而是采用小词汇量模板识别器加统计检验的方法。
- dbreunig/nesformerPython · 241 Stars
NES游戏运行、模型权重编译
查看原始资料
作者描述
An NES CPU compiled into the weights of a stock Qwen3.5 model
编报时的 AI 中文解读
这是一个将NES的6502 CPU编译进Qwen3.5模型权重的项目，通过模型预测CPU执行日志来运行NES游戏，适合对AI与模拟器交叉领域感兴趣的研究者和开发者。
- FigureSultanLast/roblox-ai-studio-agentLuau · 194 Stars
AI辅助编程、游戏开发、自动化测试
查看原始资料
作者描述
An AI agent in your terminal that builds Roblox places in Studio: parts, models, UI and Luau scripts. Place diff before every change, playtests with proof, one undo per turn.
编报时的 AI 中文解读
Stud 是一个在终端中运行的 AI 代理，用于在 Roblox Studio 中构建游戏场景，包括部件、模型、UI 和 Luau 脚本。它会在应用更改前显示差异，并自动进行游戏测试以验证修改。适用于 Roblox 开发者，尤其是使用 Rojo 或 git 的项目。
- op7418/dsh-guizang-designHTML · 191 Stars
网页PPT制作、图文卡片制作、文本润色
查看原始资料
作者描述
歸藏设计 · DeepSeek Harness 插件：网页 PPT、图文卡片、网页排版、去 AI 味四个 skill，一套设计系统；作品库、内容生产工作流、语气档案
编报时的 AI 中文解读
归藏设计是DeepSeek Harness插件，提供五个skill，用于网页PPT、图文卡片、网页排版、去AI味和语气学习，适合内容创作者制作演示、社交媒体卡片等。
- Straitbrashadow/minecraft-ai-mod-builderJava · 180 Stars
游戏内建造
查看原始资料
作者描述
Type what you want built, see it as a blueprint in your Minecraft world, place it in one step and undo it in one. Fabric companion, Paper plugin, free local models, one Windows build.
编报时的 AI 中文解读
Trowel 是一个 Minecraft 辅助工具，用户通过终端输入文字描述，即可在游戏中生成建筑蓝图，确认后一键放置，并支持一键撤销。它适用于希望在 Minecraft 世界中快速建造的玩家，支持单人和多人服务器。
- Blazenumenhir/scam-story-checkerJavaScript · 156 Stars
信息核查、证据保存、线索追踪
查看原始资料
作者描述
Does their story hold up? Paste what a seller, recruiter or landlord told you and check it against public records in thirty seconds, with proof saved for every reason.
编报时的 AI 中文解读
本地运行的免费工具，用于在交易、求职或租房前快速核查对方说辞与公开记录是否一致，并保存证据。适合个人在付款、签约前自行核实对方身份。
- Signcluengine/onchain-money-trail-analyzerJavaScript · 147 Stars
资金追踪、报告生成
查看原始资料
作者描述
Follow stolen crypto hop by hop across Bitcoin, Ethereum, TRON, BNB Chain and Solana. Labels where the money landed, keeps proof for every clue, writes the report for the police.
编报时的 AI 中文解读
追踪被盗加密货币流向的桌面应用，支持多条链，生成警方可用的报告。 免费、无需账户，提供线索保存和监控功能，适合受害者自行调查。
- DolphinStronghold9/ai-minecraft-modmakerJava · 147 Stars
游戏内建造、数据包生成、材料清单
查看原始资料
作者描述
Describe a mod or a structure and this terminal agent builds it in Minecraft: ghost preview before placing, one-step undo, datapacks that reload clean, survival materials list.
编报时的 AI 中文解读
Mortar 是一个终端 AI 代理，通过自然语言描述在 Minecraft 中生成建筑，提供幽灵预览、放置确认、一步撤销和材料清单，并支持生成数据包。适用于 Java 版和 Horizon 玩家，尤其是生存模式玩家。
- ArcSwordKiln/osint-people-checkerJavaScript · 139 Stars
背景核查、自我检查、监控提醒
查看原始资料
作者描述
Check a seller, recruiter or landlord against 3,400+ public sources before you trust them. A verdict in plain words and a saved receipt behind every finding. Runs on your PC.
编报时的 AI 中文解读
Marlowe 是一款在 Windows 上运行的免费桌面应用，用于在交易、求职或租房前核查对方背景。输入昵称、邮箱、电话或链接，它会检索 3400 多个公开来源，给出“看起来没问题”、“小心”或“有危险信号”的结论，并附上可验证的证据存档。
- DiplomatHearth/roblox-ai-place-diffRust · 138 Stars
AI辅助编程、模型切换
查看原始资料
作者描述
Every change this Roblox Studio agent makes is a place diff you approve first. Rewind any turn with one Ctrl+Z, playtests run on a copy so live data never moves.
编报时的 AI 中文解读
Baseplate 是一个在终端中运行的 AI 代理，用于在 Roblox Studio 中构建和修改游戏。它通过 MCP 服务器与 Studio 直接交互，在应用更改前显示差异，并自动进行试玩测试以验证修改。
- oilcorporatebarn/aegis-skinfeesTypeScript · 136 Stars
费用计算
查看原始资料
作者描述
Aegis SkinFees: console-grade workspace for CS2 fee math — venue models, FX conversion and net-profit estimates fully on device.
编报时的 AI 中文解读
Aegis SkinFees 是一款面向 CS2 皮肤交易者的本地优先桌面应用，提供跨平台费用计算、汇率转换和净利润估算，并集成了 Steam 市场数据获取和 Steam Guard 认证码生成。
- test-bit-sys/algeria-geoPython · 135 Stars
地理数据查询、模糊搜索、数据导出
查看原始资料
作者描述
Algeria's 69 wilayas, 538 dairas and 1,541 communes with Arabic/French search and historical administrative boundaries. Python and TypeScript libraries.
编报时的 AI 中文解读
提供阿尔及利亚行政区域（省、区、市）的权威地理数据集，包含阿拉伯语、法语和英语名称，支持模糊搜索和历史行政区划查询。适合需要处理阿尔及利亚地理数据的开发者和数据科学家。
- tapoooooty/hypergradPython · 132 Stars
永续合约交易、做空与杠杆
查看原始资料
作者描述
HIP-3 perp DEX on Hyperliquid for Solana memecoins graduating from pump.fun. GRAD index perp from $500K, own perps for coins that hold.
编报时的 AI 中文解读
HyperGrad 是一个基于 Hyperliquid HIP-3 的永续合约去中心化交易所，专注于为从 pump.fun 毕业的 Solana memecoin 提供交易市场。它包含指数合约、单币种合约、预言机、做市商和仪表盘等组件，旨在让用户能够做空和杠杆交易这些 memecoin。
- DolphinStronghold9/stolen-crypto-tracerJavaScript · 126 Stars
资金追踪、报告生成、诈骗识别
查看原始资料
作者描述
Paste a wallet or a transaction hash and follow stolen crypto to the exchange that can freeze it. Works across Bitcoin, Ethereum, TRON, BNB Chain and Solana. One-click police report.
编报时的 AI 中文解读
Wake是一款桌面应用，用于追踪被盗加密货币的流向，从钱包地址或交易哈希出发，跨链追踪资金最终到达的交易所，并一键生成警方报告。适用于被盗受害者、记者等需要追踪资金流向的人群。
- solderdatabraid29/aegis-vaultgridTypeScript · 124 Stars
交易辅助、库存管理、费用计算
查看原始资料
作者描述
Aegis VaultGrid: pairs a CS2 inventory vault with a fee-aware order planner. Local JSON state, no account needed to browse.
编报时的 AI 中文解读
Aegis VaultGrid 是一款本地优先的 CS2 交易工作区，用于查看市场机会、管理库存、制定购买计划并计算费用后的净收益。适合希望清晰掌握交易数据的玩家。
- Lavagroabide/aegis-spreadwatchTypeScript · 124 Stars
交易机会比较、净收益计算
查看原始资料
作者描述
Aegis SpreadWatch: keeps CS2 buy and sell opportunities legible — net profit, ROI and fee assumptions per venue, computed on device.
编报时的 AI 中文解读
Aegis SpreadWatch 是一款面向 CS2 交易者的本地优先桌面应用，用于比较不同交易平台的买卖机会，计算净收益、ROI 和费用假设，帮助用户做出交易决策。
其余 13 个项目见收录列表。