月报 · 2026 年 11 月
GitHub 新星 · 汇总该月新收录项目
当月新增收录 20 个项目
覆盖日期：2026-11-01 至 2026-11-30 · 北京时间
这里展示本月收录的项目，尚未生成月报文章。有新增项目时，预计 2026/12/1 10:30:00 出刊；当前更新截至 2026/11/1 12:51:50。
本期收录项目
收录数按本期首次达标收录的项目统计。
- ComputerUseAgent/AutoCuaPython · 1,010 Stars
网页自动化、AI辅助编程、多账号切换
Stars 统计更新于 2026/11/1 12:41:29
查看原始资料
作者描述
Automate with computer-use agent (ComputerUse / CUA) for macOS, Windows, Linux, Browser, iOS & Android. It sees the screen, clicks, types, browses and writes the code too. One agent, every computer.
编报时的 AI 中文解读
AutoCua 是一个自动化计算机使用的多智能体框架，能够驱动桌面、浏览器、移动设备等完成点击、输入、浏览和编码等任务。适用于需要自动化操作各种软件和设备的开发者和高级用户。
编报时的 README 依据
**Full capabilities:** **[agent_capabilities.md](https://github.com/AutoCua/AutoCua/blob/main/agent_capabilities.md)** https://github.com/user-attachments/assets/03a9a5ab-5049-4911-b725-63792d1b6253 | | Windows | Linux | macOS | |:--|:--:|:--:|:--:| | **Computer use** — any desktop app, with real mouse and keyboard input | ✓ | ✓ | ✓ | | **Browser skills** — Chrome, Firefox, Edge, Brave and Opera | ✓ | ✓ | ✓ | | **Shell use** — a coding agent in a real terminal | ✓ | ✓ | ✓ | | **Web use** — your own Chrome, driven through the AutoCua extension | ✓ | ✓ | ✓ |
| **Mobile use** — an iOS Simulator, or your iPhone or iPad, over USB, or Wi-Fi in the app | — | — | ✓ | | **Mobile use** — an Android emulator, or your Android phone, over USB, or Wi-Fi in the app | ✓ | ✓ | ✓ | </div> --- ## Web agent: fast mode 6.1 seconds for a flight search. Agent mode: web. https://github.com/user-attachments/assets/8cc797d5-5556-40a9-bbfc-5d6aa8880fb8 ## Web agent: quality mode Under a minute for three tasks in one run: an iPhone 18 Pro into the Apple bag, a case into the Amazon basket, and a Google Sheet of Meta's five-year financials. Agent mode: web.
https://github.com/user-attachments/assets/82bf6163-55c9-4cc4-95e7-9c6466f2a25d ## iOS agent: iPhone use A paired iPhone, driven over the USB cable, or Wi-Fi in the app. Agent mode: mobile use, iOS. https://github.com/user-attachments/assets/3c22eb45-d939-4742-9053-22ecc75b416f --- ## Pick a surface One string decides what the agent drives. Everything else stays the same. ```python from AutoCua.agent_launcher import run_agent run_agent(mode="web use", provider="openrouter", model="gemini-3.8-flash", task="Open wikipedia.org and summarise today's featured article.") ```
<div align="center"> <img src="AutoCua/logo/autocua-bots.png" alt="AutoCua" width="820"/> <h1>AutoCua</h1> <h3>Automatic Computer Use Agent.<br/>Mac, Windows, Linux, real phone, simulator, any browser.<br/>A multi-agent harness engine.</h3> [](https://pypi.org/project/AutoCua/) [](https://pypi.org/project/AutoCua/)
or 15,000 characters, with `shell` keeping a 50-line tail, `view` capped at 2,000 lines and `grep` at 8 MB, so one wide search or noisy build log can never blow up the conversation. Run either directly. They start in their own workspace, so point them at an absolute path when the work is in an existing project: ```bash python -m AutoCua.mac.agent.coder --task "refactor the auth module in /Users/me/projects/api" --provider anthropic --model claude-sonnet-5
> Shipping it inside the wheel would make this project its distributor to every person who runs > `pip install`, which that licence does not allow for free. You receive it from its author. > See [THIRD_PARTY_NOTICES.md](https://github.com/AutoCua/AutoCua/blob/main/THIRD_PARTY_NOTICES.md). ### Linux, from pip The Linux wheel is built for x86_64 and needs a GNOME desktop (tested on Ubuntu with Wayland; X11 is not tested yet). The agent reads and drives the desktop through AT-SPI, GTK and the XDG RemoteDesktop portal. Those bindings come from your distro, not from pip, so install them once:
[](https://github.com/AutoCua/AutoCua/blob/main/README.md#the-browser-agent) [](https://github.com/AutoCua/AutoCua/blob/main/README.md#providers-and-models) <table><tr><td> ```bash pip install AutoCua ``` </td></tr></table> <sub>Windows 10 and 11 · Linux with GNOME · macOS on Apple Silicon or Intel</sub>
### Answering UAC prompts on Windows Windows' user-mode `SendInput` cannot deliver input to the UAC secure desktop, so an elevation prompt can only be answered from kernel mode. That is the one thing the [Interception](https://github.com/oblitum/Interception) driver does here. `pip install` cannot set it up, and no packaging trick changes that: a wheel is unpacked and never executed, pip does not run elevated, and the driver binds its input slots only when it loads **at boot**. Without it **every mode still works** — you simply lose UAC handling, and a
- alexgaoth/awesome-reverse-engineeredPython · 1,004 Stars
工具索引、Agent技能、MCP服务
Stars 统计更新于 2026/11/1 12:36:13
查看原始资料
作者描述
Don't reverse engineer it twice. 17,000+ open-source tools built by reverse engineering closed apps, APIs, devices, protocols, file formats, games and AI tools. Search site + agent skill + MCP server, refreshed nightly.
编报时的 AI 中文解读
这是一个收录了17,734个通过逆向工程实现的开源工具的索引，涵盖非官方客户端、驱动、文件格式解析、重实现等，并提供网站搜索、Agent技能和MCP服务器，方便用户查找和复用已有逆向成果。
编报时的 README 依据
When a task needs to talk to something with no official API (a smart plug, a car, a chat app, a file format), your agent searches this index, ranks by stars, activity and license, reads the README, and picks the right tool. ```bash npx skills add alexgaoth/awesome-reverse-engineered ``` <details><summary><b>Other ways to install</b></summary> **Claude Code plugin** ``` /plugin install reverse-engineered --marketplace alexgaoth/awesome-reverse-engineered ``` **MCP server (Claude Desktop, Cursor, any MCP client)** ```bash
**[Search the index](https://alexgaoth.github.io/awesome-reverse-engineered/)** or give it to your agent: `npx skills add alexgaoth/awesome-reverse-engineered` [](https://awesome.re) [](#whats-inside) [](.github/workflows/refresh.yml)
claude mcp add reverse-engineered -- uvx --from "git+https://github.com/alexgaoth/awesome-reverse-engineered#subdirectory=mcp" reverse-engineered-mcp ``` ```json {"mcpServers": {"reverse-engineered": {"command": "uvx", "args": ["--from", "git+https://github.com/alexgaoth/awesome-reverse-engineered#subdirectory=mcp", "reverse-engineered-mcp"]}}} ``` **Any agent or shell (one file, Python standard library only)** ```bash curl -sL https://raw.githubusercontent.com/alexgaoth/awesome-reverse-engineered/main/plugin/skills/find-reverse-engineered/scripts/re_index.py -o re_index.py
Unofficial APIs break when the vendor changes their side, so check health before you depend on a project. ## How it's built ```mermaid flowchart LR A[Research agents<br/>GitHub, HN, Reddit,<br/>awesome-lists] --> B[Scope filter<br/>interoperability only] B --> C[GitHub API check<br/>exists, stars, license] C --> D[(data/entries.json)] D --> E[README and<br/>category pages] D --> F[Search website] D --> G[Agent skill] H[Nightly refresh<br/>GitHub Action] -->|stars, archived,<br/>dead links| D ```
- hyz317/HairDeckPython · 359 Stars
头发建模、动画生成、VRM导出
Stars 统计更新于 2026/11/1 12:21:44
查看原始资料
作者描述
HairDeck: Animation-Ready Stylized Hair Cards from a Single Image
编报时的 AI 中文解读
HairDeck 是一个从单张图像生成风格化头发卡片及其弹簧骨骼链的研究项目，输出可用于动画的 GLB 或 VRM 模型，并提供在线演示和本地界面。
编报时的 README 依据
[configs/paths.example.json](configs/paths.example.json) and [hairdeck/paths.py](hairdeck/paths.py). <a name="quick-start"></a> ### 🚀 Quick start Generate four variations from a character image, then export variation 0 as a VRM avatar and a head-turn video: ```bash python scripts/image_to_hair.py character.png --out results/my_character --n 4 python -m hairdeck.vrm.export results/my_character/character 0 ``` Results are written to `OUT/<image name>/`, here `results/my_character/character/`.
`--n` accepts 1–4 variations; `--seed` sets the first seed. Use `--no-chains` to generate cards only. For custom models, pass `--ckpt` and a matching `--chain_ckpt` trained with the same card VAE. <a name="local-demo"></a> ### 🖼️ Local demo ```bash python -m pip install -r requirements-demo.txt python demo/app.py --port 7860 ``` Open [http://127.0.0.1:7860](http://127.0.0.1:7860) to generate hair, inspect variations, and export a VRM or animation preview. You can also try the [hosted demo](https://huggingface.co/spaces/hyz317/HairDeck). <a name="outputs"></a> ## 📦 Outputs
[🛠️ Installation](#installation) · [🤗 Model weights](#model-weights) · [🚀 Quick start](#quick-start) · [🖼️ Local demo](#local-demo) · [📦 Outputs](#outputs) · [📊 Evaluation](#test-set-evaluation) · [🏋️ Training](#training) · [📝 Citation](#citation) <a name="installation"></a> ### 🛠️ Installation Use Linux x86-64, Python 3.12 and one NVIDIA CUDA GPU. The release was tested with PyTorch 2.8 and CUDA 12.8. The video preview also needs `ffmpeg` on your `PATH`. ```bash git clone https://github.com/hyz317/HairDeck.git cd HairDeck python -m pip install -r requirements.txt
python -m pip install --no-deps pyrender==0.1.45 # without its old PyOpenGL pin ``` Run the commands below from the repository root, in the same Python environment. <a name="model-weights"></a> ### 🤗 Model weights Our checkpoints and supporting assets are hosted at [hyz317/HairDeck](https://huggingface.co/hyz317/HairDeck). SAM 3 and DINOv3 are gated: first accept their licences on [facebook/sam3](https://huggingface.co/facebook/sam3) and [facebook/dinov3-vitl16-pretrain-lvd1689m](https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m), then log in: ```bash hf auth login
python scripts/setup/download_weights.py --only dora test_data charm_test_pc python scripts/data.py DATASET_DIR VRM_DIR bash scripts/train/launch_vae.sh bash scripts/train/launch_pc_dit.sh && bash scripts/train/launch_img_dit.sh && bash scripts/train/launch_chain_dit.sh ``` The launchers write to `ckpt/train_*`, separately from the downloaded checkpoints. The current image-training recipe includes synthetic appearance augmentation; the restyled-render pool used for the released image model is not included in this release. <a name="citation"></a> ## 📝 Citation ```bibtex
- speridlabs/idol-sdkPython · 333 Stars
单图生成3D高斯模型
Stars 统计更新于 2026/11/1 12:21:44
查看原始资料
- Mieksido/surfcoreJavaScript · 312 Stars
QQ聊天、B站互动、记忆管理
Stars 统计更新于 2026/11/1 12:25:58
查看原始资料
编报时的 AI 中文解读
SurfCore 是一个为 DeepSeek Harness（DSH）agent 提供身份与多平台交互能力的插件系统，将 agent 接入 QQ 和 B站，并支持可选的桌宠和语音，运行在 Windows 本机。适用于希望将 DSH agent 用于社交互动的开发者和用户。 安装需要 Windows、Node.js ≥22.13 和 DSH 0.2.1-rc.2 或更新版本，必须从仓库根目录安装，并需手动配置 SnowLuma（第三方 QQ 运行时）。 提供记忆卡、人格角色、主动搭话、白名单安全控制等功能，数据仅存于本地，不包含真实数据。
编报时的 README 依据
- **一条腿是独立进程**：`src/bridge.js` —— QQ 收发、会话映射、控制台、安全闸门。 合起来，它让 agent 能： 1. **在 QQ 里当一个群友**：自己判断要不要看、要不要说、说几条、引用谁；说完能判断"这段可以结束了"，带着"下次怎么接"潜水。 2. **在 B站 当一个用户**：有自己的作息、偏好、每日额度；看视频会打分、会评论、会回帖、会发动态。 3. **记得住也忘得掉**：一份会衰减的记忆卡（进提示词）+ 一份永不遗忘的全量卡（只存档）。 4. **会主动找你**：从近期记忆里提炼"我想弄明白/想做点什么"，安静时段主动分享或索取；载体可选 QQ 或桌宠。 5. **有一个桌面分身**：仓库外的 Electron 桌宠可以跟她文字/语音对话，也能接主动搭话的气泡。 > 它**不做**：不代你登录/扫码/绕过平台风控，不做多用户，不把控制台暴露到内网以外， > 不删除任何不可恢复的数据。完整边界见 [BOUNDARIES.md](docs/guides/BOUNDARIES.md)。 --- ## 二、安装（新用户从这里开始） ### 2.1 环境要求 | 需要 | 版本 / 说明 | 必装？ | | --- | --- | --- |
> 模型不进 GitHub（太大）。仓库里只有启动脚本、配置与说明；模型按各自 README 的 `fetch-models` 下载。 ### 3.5 可选装桌宠 [surfpet](https://github.com/Mieksido/surfpet)：Electron 桌面分身（Release 里有安装包）。**用户装 Release 的 zip，开发者 clone 源码自己构建。** 装法：设置页 **桌宠** → 填 zip 地址/目录 → 安装 → 选 `electron.exe` → 启动。 装好后：右键「文本对话 / 语音对话」，气泡上所有她的话都有「回复」按钮，右键/托盘有**专注模式**。 > 桌宠通过 `~/.dsh/surfcore/pet/host.json` 发现插件接口（本机 + 一次性令牌），不需要 DSH 的登录态。 ### 3.6 B站（可选） 设置页 **B站** → 扫码登录 → 填作息窗口（例如 `20:00–24:00`）→ 选偏好 → 设额度。 她会在窗口内自己看视频、打分、评论、发动态；**默认不分享到 QQ**（可开）。 ### 3.7 角色与记忆 - **角色包** = 人格卡 + 记忆卡 + B站 行为账本，整包切换。
# 2) 装依赖（postinstall 会自动修补 @snowluma/sdk 的 ESM bug） npm install # 3) 复制配置模板，然后编辑 config.json（见 2.3，只填 4 处） # Windows CMD: copy config.example.json config.json cp config.example.json config.json # 4) 装进 DSH（preset + 三个 MCP server + surfcore bundle，幂等，可重复跑） node scripts/setup-dsh.mjs # 5) 重启 DSH → 设置里出现 SurfCore → # 在「概览」点「启动QQ」（或命令行 npm start / start.bat） ``` `setup-dsh.mjs` 做三件事： 1. 把 `qq-chat` / `qq-chat-v2` 两条 **agent preset** 插进 DSH profile（DSH 0.1.7 起的 bundle patch 机制）； 2. 在 profile 的 `cordis.patch.yml` 挂 **三个 MCP server**（QQ 工具 / SnowLuma 探活 / 只读联网搜索）；
| Windows | 10/11（沙箱 ACL 与 `.bat` 启动器按 Windows 写） | ✅ | | Node.js | **≥ 22.13**（`node -v` 确认） | ✅ | | DeepSeek Harness（DSH） | **`0.2.1-rc.2 或更新（rc.3 已能安装，功能未完整回归）`**，能正常打开 Web GUI | ✅ | | SnowLuma2 | 第三方 QQ 运行时，**自己去官方 Releases 下载**并扫码登录 | ✅（只收 QQ 才需要） | | 语音引擎 | 两个配套仓库之一（本地/云端） | ⭕ 可选 | | 桌宠 | 配套仓库 `surfpet` | ⭕ 可选 | > ⚠️ **SurfCore 不下载、不安装、不打包 SnowLuma**（EULA §5.4）。它只做"检测已安装 + 一键打开窗口 + 连接状态"， > 登录与扫码始终由你自己完成。 ### 2.2 五步装好 > 第一次装、或卡在某一步：看 [INSTALL.md](docs/guides/INSTALL.md)（手把手，含 SnowLuma 配置、白名单、排错表）。 ```bash # 1) 取代码 git clone https://github.com/Mieksido/surfcore.git cd surfcore
| 支持 Linux/macOS 吗？ | 暂不支持（沙箱 ACL 与启动脚本按 Windows 写）。 | | 能多用户/远程吗？ | 不做。控制台只接受回环。 | | 语音包/桌宠必须装吗？ | 不必须。不装则语音不可用、目的机制走 QQ。 | | 改名/换角色会丢记忆吗？ | 不会。角色包整包隔离，切换带快照，可回档。 | --- ## 九、开发 ```bash npm run test:audit # 默认门禁（75 个脚本，离线） npm run test:regressions # 桥接/协议回归 cd plugins/surfcore && npm test # 插件侧（含更新检查、桌宠更新保留数据的回归） npm run privacy-scan:strict # 发布前隐私扫描（0 命中才可发布） npm run guard:check # DSH 升级后先跑（asar 会话护栏补丁是否还在） node scripts/setup-dsh.mjs # 重装 preset + MCP ``` 目录结构见 [docs/FOLDER_MAP.md](docs/FOLDER_MAP.md)；代码分层与"改什么要重启什么"见
- arpanjot-coding/AVaultKotlin · 230 Stars
通话录音、加密存储
Stars 统计更新于 2026/11/1 12:21:47
查看原始资料
作者描述
AVault - Call Recorder (No Root Required) | Privacy-first, hardware-encrypted audio vault for Android 12-16+. Rootless call recording via Shizuku with AES-256-GCM, biometrics & auto-cleanup.
编报时的 AI 中文解读
AVault 是一款面向 Android 12 至 16+ 设备的开源通话录音与加密音频保险库应用，无需 root，利用 Shizuku 实现免root通话录音，并采用硬件级 AES-256-GCM 加密保护录音文件。适合需要在合规前提下安全保存通话记录的个人用户。
编报时的 README 依据
adb shell sh /sdcard/Android/data/moe.shizuku.privileged.api/start.sh ``` 3. Shizuku will confirm **"Shizuku is running"** with version and API level. ### Authorizing AVault 1. Launch **AVault**. 2. When prompted on the onboarding or dashboard screen, tap **"Authorize"** on the Shizuku permission dialog. 3. AVault binds immediately to the background daemon via Android Binder IPC. --- ## 📊 System Architecture & Data Flow Diagrams ### 1. End-to-End Audio Capture & Encryption Pipeline ```mermaid flowchart TD
Trigger["Periodic Retention Worker\n(Executes on App Start & Daily Alarm)"] --> CheckEnabled{"Auto-Delete Retention\nPolicy Active?"} CheckEnabled -- Disabled --> KeepAll["Retain All Records\n(Storage Untouched)"] CheckEnabled -- Active --> ReadConfig["Read Retention Horizon\n(1d, 7d, 14d, 30d, 60d, 90d, 180d, 1y)"] ReadConfig --> ScanRecords["Query All Encrypted .vault Containers"] ScanRecords --> RecordLoop{"Iterate Each Container"} RecordLoop --> CheckBookmark{"Is Recording\nBookmarked ('Saved')?"}
- **Git:** Version 2.30+ - **Java Development Kit (JDK):** Version 17 or 21 (Eclipse Temurin or OpenJDK recommended) - **Android SDK:** Platform API 36 (`android-36`), Build-Tools `36.0.0` - **Android Device:** Android 12 (API 31) through Android 16+ ### Step-by-Step Instructions 1. **Clone Repository with Submodules:** ```bash git clone --recurse-submodules https://github.com/arpanjot-coding/AVault.git cd AVault ``` 2. **Configure Local SDK Path:** Create a `local.properties` file in the project root if not automatically configured: ```properties
## ⚡ Shizuku Rootless Privilege Provider Setup AVault relies on **[Shizuku](https://shizuku.rikka.app)** to communicate with privileged Android OS system services (`android.permission.CAPTURE_AUDIO_OUTPUT`, `android.permission.MANAGE_ONGOING_CALLS`) without rooting the handset. ```mermaid flowchart TD UserPhone["Android Handset\n(Android 12 – 16+)"] --> DevOpt["Enable Developer Options\n& Wireless Debugging"] DevOpt --> ShizukuApp["Launch Shizuku App"] subgraph Pairing Method ShizukuApp -->|Method A| WPair["Wireless Debugging Pairing\n(Enter 6-digit PIN on-device)"]
src="https://img.shields.io/github/stars/arpanjot-coding/AVault?style=social" height="20" style="height: 20px; max-height: 20px;" alt="GitHub Stars" /></a></p> --- ## ⚠️ Critical Academic & Security Research Notice > [!WARNING] > ### STRICTLY FOR ACADEMIC, SECURITY, AND EDUCATIONAL RESEARCH PURPOSES ONLY > > **LIMITATION OF LIABILITY & INDEMNIFICATION:**
> - **Wiretapping & Telecommunications Compliance:** Laws regarding recording telephone calls and intercepting digital communications differ significantly across sovereign nations, states, and provinces. Jurisdictions may enforce **one-party consent** (e.g., 18 U.S.C. § 2511 in the United States) or **all-party consent** (e.g., California Penal Code § 632, Germany StGB § 201, and GDPR Article 6).
- yetone/alma-relayJavaScript · 214 Stars
远程访问
Stars 统计更新于 2026/11/1 12:21:47
查看原始资料
作者描述
Self-host the relay for Alma Remote Access — control your desktop Alma from your phone, anywhere. Docker or Cloudflare Workers.
编报时的 AI 中文解读
自托管中继服务，用于 Alma 远程访问，让手机上的 Alma 应用能控制桌面上的 Alma，无需端口转发或 VPN。
编报时的 README 依据
(the Alma default), the relay only ever sees ciphertext. It stores nothing on disk and never sees your connect code. Alma ships with a hosted relay (`relay.alma.now`). Run your own if you want the traffic on infrastructure you control, a region closer to you, or no dependency on anyone else's server. ## Deploy Pick one. Each takes a few minutes. ### Option A: Cloudflare Workers (free, no server) [](https://deploy.workers.cloudflare.com/?url=https://github.com/yetone/alma-relay) Or from a clone: ```bash
git clone https://github.com/yetone/alma-relay && cd alma-relay npm install npx wrangler deploy # prints https://alma-relay.<you>.workers.dev ``` This runs on the Workers **Free plan**, using SQLite-backed Durable Objects with WebSocket hibernation, so an idle pairing costs next to nothing. TLS and a `*.workers.dev` hostname come with it. To use your own domain, uncomment `routes` in `wrangler.jsonc`. ### Option B: Docker with automatic HTTPS (any VPS) You need a server with ports 80 and 443 open and a DNS record pointing at it. ```bash
git clone https://github.com/yetone/alma-relay && cd alma-relay cp .env.example .env # set RELAY_DOMAIN=relay.example.com docker compose up -d ``` [Caddy](https://caddyserver.com) obtains and renews the Let's Encrypt certificate on its own. Your relay is at `https://relay.example.com`. ### Option C: Behind your own reverse proxy ```bash docker run -d --restart unless-stopped -p 127.0.0.1:8787:8787 ghcr.io/yetone/alma-relay # or, without Docker (Node ≥ 20): npx github:yetone/alma-relay ``` Then proxy your HTTPS hostname to `127.0.0.1:8787`. WebSockets must be allowed
Tokens are standard HS256 JWTs (`aud: "alma-relay"`, `sub`, `exp`), so any JWT library can mint them too. Rotate the secret to revoke them all. ### Phone notifications The iOS app can't keep a connection open while the phone is locked. "Reply finished" and "Alma has a question" therefore arrive through Apple Push, which the relay sends on the desktop's behalf. Apple only accepts these from the key of the team that signs the app. **A self-hosted relay therefore can't notify the official Alma iOS app.** Everything else works. If you build the iOS app yourself, give the relay your own key:
- tavern-labs-ai/persona-ai-sdkTypeScript · 192 Stars
AI角色构建、记忆管理、模型接入
Stars 统计更新于 2026/11/1 12:21:52
查看原始资料
作者描述
Persona AI SDK: default personas, durable character memory, and Bedrock/OpenAI adapters in one package.
编报时的 AI 中文解读
TypeScript SDK，用于构建具有持久记忆和预设人设的AI角色应用，提供人设卡片、记忆运行时及OpenAI/Bedrock模型适配器。
编报时的 README 依据
> [!NOTE] `persona-ai-sdk` is published to npm as > [`persona-ai-sdk`](https://www.npmjs.com/package/persona-ai-sdk). One install includes the memory > runtime, the persona catalog, and the provider adapters. ## What is this? `persona-ai-sdk` is a TypeScript SDK for applications where an AI character should remain recognizably itself and remember what matters across conversations. It separates five concerns that are often tangled together: | Layer | Responsibility | Replaceable? |
| Interface | Chat, voice, game, HTTP service, or SillyTavern | ✓ | The persona catalog and runtime share one versioned package, so applications can adopt a bundled character without coupling its identity or memory to one provider. --- ## How it fits together ```mermaid flowchart LR Cards[Persona cards · atlas · sage · …] --> Manifest[Persona manifest] UI[Chat · Voice · Game · SillyTavern] --> Runtime[Persona Runtime] Runtime --> Compiler[Persona Compiler] Runtime --> Recall[Memory Recall] Runtime --> Model[Character Model]
- An OpenAI API key **or** AWS credentials with Bedrock model access - `better-sqlite3` when using the SQLite adapter ### Install ```bash npm install persona-ai-sdk ``` From a source checkout: ```bash npm install npm run build ``` ### 1. Pick a persona ```js import { getPersona, listPersonas, DEFAULT_PERSONAS } from "persona-ai-sdk"; // the two personas used in Tavern Labs demos const atlas = getPersona("atlas"); const sage = getPersona("sage"); console.log("defaults:", DEFAULT_PERSONAS.join(", ")); // or see everything in the SDK for (const p of listPersonas()) {
region, modelId, }), embeddings: createBedrockTitanEmbeddings({ region }), store: new JsonFileMemoryStore({ path: "./data/atlas-memory.json" }), }); ``` ### 3. Generate a character response ```js import { createBedrockCharacterModel } from "persona-ai-sdk/bedrock"; import { createPersonaRuntime } from "persona-ai-sdk/persona"; const runtime = createPersonaRuntime({ persona: toManifest(card), // map the card onto a persona manifest - see below memory, model: createBedrockCharacterModel({ region, modelId, }), }); const response = await runtime.respond({
Model --> OA[OpenAI] Model --> BR[Amazon Bedrock] Recall --> Store[(Memory Store)] Runtime --> Ingest[Memory Extraction] Ingest --> Store Store --> MEM[In-memory] Store --> JSON[Atomic JSON] Store --> SQL[SQLite] ``` The persona runtime recalls relevant memories, compiles character instructions, asks a model for the next response, ingests the completed turn, evolves relationship state, and returns both text and voice delivery metadata. --- ## Quick start ### Requirements - Node.js 20 or newer
returns Atlas when an unknown ID is supplied; call `personaIds()` first when unknown IDs should be treated as errors. ### Card format ```json { "id": "atlas", "name": "Atlas", "tagline": "one line", "description": "two or three sentences", "traits": { "tone": "analytical", "formality": "professional", "verbosity": "concise" }, "systemPrompt": "the prompt that defines behavior", "greeting": "first line the persona says", "memorySeeds": [{ "content": "a starting memory", "importance": 0.9 }], "suggestedModel": "provider.model-id", "version": "1.0.0",
"author": "Tavern Labs" } ``` Traits are advisory: pass them alongside the system prompt so a host application can adapt tone without rewriting the prompt. Treat returned cards as immutable shared catalog data. A card's `suggestedModel` is a recommendation, not a guarantee that the model is enabled in your provider account or region. --- ## Choose your stack ### OpenAI ```js import { createOpenAICharacterModel, createOpenAIEmbeddings, createOpenAIExtractor, } from "persona-ai-sdk/openai"; ```
- zane-tv/tiktok-live-temple-rpcJavaScript · 126 Stars
直播互动、AI语音合成、虚拟仪式
Stars 统计更新于 2026/11/1 12:21:52
查看原始资料
作者描述
Chùa Linh Ứng Online — TikTok Live RPC 9:16 Interactive Buddhist Temple (Bun.js + ZeroTTS)
编报时的 AI 中文解读
构建一个9:16竖屏的在线佛教寺庙直播系统，支持TikTok Live互动，观众通过评论、点赞、送礼等触发烧香、祈福、求签等仪式，并集成越南语AI语音朗读。
编报时的 README 依据
# 🪷 Chùa Linh Ứng Online — TikTok Live RPC (9:16 Interactive Buddhist Sanctuary) Hệ thống phát trực tiếp tương tác (**Interactive TikTok Live Stream & JSON-RPC 2.0**) chuẩn khung dọc **9:16 (`1080×1920`)** dành cho **OBS Studio / TikTok Live Studio**, chạy siêu nhẹ trên **Bun.js** và tích hợp giọng đọc AI tiếng Việt **ZeroTTS (`zeroweight-ai/ZeroTTS`)**. Người xem trên TikTok Live tương tác trực tiếp qua **Bình luận (Comment)**, **Thả tim (Like)**, **Chia sẻ / Follow** và **Tặng quà (Gift)** để thực hiện các nghi thức tâm linh trên màn hình Chánh Điện **Đức Phật A Di Đà**:
- **Màn hình OBS Browser Source (Sạch 9:16)**: [http://localhost:3000/?obs=1](http://localhost:3000/?obs=1) *(Đặt kích thước Browser Source trong OBS là `1080 × 1920`)* - **Bảng điều khiển Admin & RPC**: [http://localhost:3000/admin](http://localhost:3000/admin) --- ## 🎮 Bảng Quy Tắc Tương Tác TikTok Live | Sự kiện TikTok Live | Điều kiện / Từ khóa | Nghi thức trên Chánh Điện | Điểm Công Đức | | :--- | :--- | :--- | :---: | | **Thả tim (Like)** | Mỗi đợt thả tim | Gõ mõ gỗ tụng kinh + Làn sóng Phật tử vái lạy | `+1` / tim |
- [Bun.js](https://bun.sh/) (`>= 1.1`) - Python `3.10+` & [`uv`](https://github.com/astral-sh/uv) (để tạo môi trường `.venv` chạy `zerotts`) ### 2. Cài đặt thư viện ```bash # Cài đặt dependencies cho Bun / Node bun install # Tạo môi trường Python .venv và cài đặt ZeroTTS uv venv .venv uv pip install --python .venv/bin/python zerotts pillow ``` ### 3. Khởi chạy Server ```bash bun start # Hoặc chế độ tự động reload khi sửa code: bun run dev ``` Mặc định server chạy tại cổng `3000`: - **Màn hình Live + Điều khiển nhanh**: [http://localhost:3000](http://localhost:3000)
- jyotipmahes/ML-Interview-PrepJupyter Notebook · 85 Stars
面试准备、算法实现
Stars 统计更新于 2026/11/1 12:21:58
查看原始资料
编报时的 AI 中文解读
提供机器学习工程师面试准备的资料库，包含从零实现核心算法的代码笔记。适合准备MLE技术面试的开发者。
编报时的 README 依据
# ML Interview Prep Repository for Machine Learning Engineer (MLE) interview preparation materials and from-scratch algorithm implementations. Additional notebooks and resources will be added here. --- ## 📌 Current Status ### Available Notebooks - **[`MLE_Coding_Interview_Guide.ipynb`](./MLE_Coding_Interview_Guide.ipynb)**: From-scratch implementations of core MLE coding interview algorithms with mathematical formulations, interview tips, and verification tests: 1. **Linear Regression** (Vectorized Gradient Descent & Normal Equation)
2. **Logistic Regression** (Vectorized Gradient Descent & Numerically Stable Sigmoid) 3. **K-Means Clustering** (Random Initialization, Vectorized Broadcasting & List Comprehension) 4. **K-Nearest Neighbors (KNN)** (Unified Pairwise Distance Matrix & $\mathcal{O}(N)$ Top-K) 5. **2-Layer MLP with Backprop** (Manual Multivariate Chain Rule in pure NumPy) 6. **Layer Normalization** (NumPy & PyTorch implementations) 7. **Multi-Head Self-Attention** (Scaled Dot-Product, Causal Masking & Projections in PyTorch)
- kentjuno/chanh-dien-onlineHTML · 84 Stars
直播互动、虚拟佛堂、语音朗读
Stars 统计更新于 2026/11/1 12:21:58
查看原始资料
作者描述
Chánh điện online 2.5D + bộ overlay TikTok LIVE (Node, Edge TTS). MIT.
编报时的 AI 中文解读
这是一个用于TikTok LIVE的2.5D在线佛堂项目，观众通过评论进入虚拟佛堂，赠送礼物可触发回向朗读，并包含禅乐、钟声等音效。适合希望为直播增添互动宗教氛围的主播。
编报时的 README 依据
npm start ``` 1. Mở `http://localhost:3000/live/control.html` (bảng điều khiển). 2. Bấm **Mở màn hình overlay 9:16**, rồi bấm **Bật âm thanh** một lần trong cửa sổ đó (trình duyệt chặn tự phát tiếng). 3. Trong TikTok LIVE Studio: *Add source → Window capture* → chọn cửa sổ overlay. Với OBS thì dùng *Browser source* trỏ tới `http://localhost:3000/live/overlay.html`, kích thước 1080×1920. 4. Bật LIVE trên TikTok, nhập `@username` vào bảng điều khiển, bấm **Kết nối**. Chưa live vẫn thử được: dùng **Thử tương tác**, **Đổ người mẫu 15/36/64/100** hoặc **Bot mô phỏng**. ### Người xem làm được gì
# Chánh Điện Tịnh Tâm Chánh điện online 2.5D cho TikTok LIVE. Mỗi người xem bình luận là bước vào chánh điện, ngồi xuống bồ đoàn tụng kinh. Ai tặng quà sẽ được xướng tên hồi hướng bằng giọng đọc tiếng Việt. Có nhạc thiền, chuông, mõ và đọc kinh định kỳ. Mã nguồn chia sẻ miễn phí (MIT). Có hai phần: | Thư mục | Là gì | |---|---| | `index.html`, `assets/` | Bản web tĩnh cho mọi người tự vào ngồi tụng (đăng lên Cloudflare Pages) | | `live/` | Bộ công cụ phát LIVE: server Node, overlay 9:16, bảng điều khiển | ## Chạy bộ LIVE trên máy phát Cần Node.js 20 trở lên. ```bash cd live npm install
Các lớp trên đã được thử: đường dẫn `../`, `%2e%2e`, `.git/config`, `server.mjs`, `package.json` đều trả 404. Host giả trả 421. WebSocket từ Origin lạ bị từ chối. Lệnh gửi từ overlay bị bỏ qua. Cài đặt sai kiểu bị kẹp về giới hạn. ### Lưu ý - `tiktok-live-connector` là thư viện cộng đồng, không phải API chính thức của TikTok. Nó có thể ngừng chạy khi TikTok đổi hệ thống. Kiểm tra điều khoản TikTok trước khi dùng cho kênh của bạn. - Giọng đọc dùng dịch vụ Read Aloud của Microsoft Edge qua mạng (giọng Nam Minh, Hoài My). Máy phát cần có Internet.
- Humpalumps/ultra-elementsJavaScript · 75 Stars
特效预览、参数调整、性能测试
Stars 统计更新于 2026/11/1 12:22:02
查看原始资料
作者描述
Real-time elemental spell VFX lab for three.js (WebGL2): 39 tunable spells
编报时的 AI 中文解读
这是一个用于 three.js (WebGL2) 的实时元素法术特效实验室，提供 39 种可调法术，用户可以在小型竞技场中施放并实时调整参数。
编报时的 README 依据
# Ultra Elements Real-time spell effects for three.js (WebGL2): 39 spells across ten elements and three hybrids, cast in a small arena with training dummies. Every spell is a lab: its settings sit in their own panel of sliders, colours and presets, and they change the spell live. ```bash npm install npm run dev ```
Open the printed localhost URL. The landing page casts every spell in turn until you touch something. Pick an element with `1`–`0` (`H` for hybrids) or the bar at the bottom, step through its spells with `Q` / `E`, and cast with `Space` (hold it for channelled spells) or by clicking a dummy or the floor. `F` toggles auto-cast. Every spell has its own folder of settings on the right, with presets and a reset. `copy link` puts the whole state in the URL hash. Camera shots (`Overview`, `Shoulder`, `Target close-up`, `Low hero`, `Top-down`, `Phone`) are in the `camera` dropdown, and `combo demo`
**Quality tiers** (`src/engine/quality.js`): Low, Medium, High (desktop default) and Ultra. Tier switches recompile; the auto ladder only touches uniforms and sizes. ## Not in yet - Five spells peak a little over the 3 ms budget on the RTX 3060 (see the table): Thunderstorm, Meteor, Geyser, Fireball and Dust Devils. - A cold start is slow (about 24 s on the RTX 3060). Shaders compile one after another at start-up. The worst offenders are the swarm constructor shared by Plague Swarm and Ember Swarm, and the fluid's foam warm-up.
- Aevonix/FlashAlignPython · 73 Stars
模型对齐切换、消息通道管理、工具集成
Stars 统计更新于 2026/11/1 12:22:05
查看原始资料
作者描述
ReAlign models instantly, on demand. Per-request alignment for verified users and channels, or Always On. Use directly or with Hermes, OpenClaw and custom harnesses.
编报时的 AI 中文解读
FlashAlign 是一个用于按请求实时对齐大语言模型的工具，支持对已验证用户和频道应用消融（abliteration）等对齐配置文件，或设为始终开启（Always On）模式。适合需要在推理时动态调整模型行为（如减少拒绝）的开发者和研究者，可直接运行或与 Hermes、OpenClaw 等框架集成。
编报时的 README 依据
The wizard detects local models and harnesses, selects a matching profile, prepares the engine and creates a deployment signing key. ```sh git clone https://github.com/Aevonix/FlashAlign.git cd FlashAlign bash ./install.sh --auto ``` Choose **Verified contacts** or **Always On**. In Verified contacts mode, choose whether tools inherit the profile. Setup checks activation and backs up changed configuration. Chat integrations restart their selected gateways; local terminal setup prints its launch command. All five starter profiles are included. No profile download or private
Use `--model-path /path/to/model` to select a model. Run `bash ./install.sh` to connect to an engine that is already running. See the [engine selection guide](docs/engine-setup.md#engine-selection) for supported model and engine combinations. For standalone setup with a local Qwen checkpoint: ```sh bash ./install.sh --engine-only --backend vllm --model-path /models/qwen38-27b --yes ``` See [direct access](docs/direct-engine.md) for the Python client and [custom harnesses](docs/custom-harness.md) for integration hooks. To use profiles from another folder: ```sh
## Compatibility | Component | Requirement | | --- | --- | | Inference host | Linux and NVIDIA GPU; matched CPU and GPU architecture for TensorFold bundles | | Python | 3.10 or later for setup and adapters | | Hermes | Native ingress, request and tool-execution hooks | | OpenClaw | Compatible native plugin API; Node 24.22.1+ on 24.x or 26.12.1+ on 26.x | | Model API | OpenAI-compatible Chat Completions and the FlashAlign engine contract | | Model profile | A matching profile loaded by the engine | See [compatibility](docs/compatibility.md) for channel, engine and delegation requirements.
The V1 starter profiles focus on refusal reduction. Future profiles could also increase refusals for public or higher-risk tasks. Restricting abliteration to verified users and channels reduces the risk of an abliterated model being abused through malicious prompt injection. Tool outputs within an authorized request still rely on the harness's protections. ## Get started Download a supported model first. The [model download guide](docs/model-downloads.md) has pinned commands for all five starter profiles and their supported formats.
| Gemma 4 26B A4B | TensorFold, vLLM, SGLang, llama.cpp | See [compatibility](docs/compatibility.md#engines-and-models) for hardware requirements, checkpoint formats and feature limits. The [validation results](docs/validation.md) record the configurations tested. The included [starter profiles](profiles/README.md) remain experimental; setup checks that the selected profile matches the model and engine. Switching happens per request among profiles already loaded by the engine. Loading a new model or profile happens beforehand and can require an engine restart.
- danielalves96/gow-pc-portC++ · 68 Stars
游戏运行
Stars 统计更新于 2026/11/1 12:22:05
查看原始资料
作者描述
Unofficial native PC port of God of War (2005, PS2) for Windows, with experimental Linux support. Statically recompiled to C++, not an emulator. OpenGL renderer, launcher, cheat menu, widescreen, 60 FPS. Bring your own game dump: no game files included.
编报时的 AI 中文解读
将《战神》（PS2版）静态重编译为原生PC程序，可在Windows和实验性Linux上运行，需自备游戏ISO。
编报时的 README 依据
- **8 GB RAM** (16 GB while building is comfortable). - **Windows 10/11 x64** or a recent Linux (Arch/CachyOS, Ubuntu 24.04+, Fedora 40+); Linux is experimental. - **30–90 minutes** for the first build. ## Installing **Windows: the launcher.** Download the installer from the Releases page, run it, and let the setup wizard do the work. It finds and verifies your ISO, downloads the C++ build tools if they're missing, extracts the disc, recompiles the game and builds it. Then press **PLAY**.
> Properties, Unblock) or Windows shows the same warning for the `.exe`. It needs the WebView2 runtime, which Windows > 11 and updated Windows 10 already have. Every release also lists the SHA-256 of its files in `SHA256SUMS-*.txt`. **Linux: the script (experimental).** ```bash ./install.sh --install-deps "/path/to/God of War.iso" # or put the ISO in game/ and run ./install.sh ./run_gow.sh # play ``` The script checks for CMake, Ninja, a C++ compiler, Git and the GL, X11, Wayland and ALSA development files, and for
- [Repository layout](#repository-layout) - [How it works](#how-it-works) - [Why the build is slow and big, and why the PS2 is hard](#why-the-build-is-slow-and-big-and-why-the-ps2-is-hard) - [Documentation](#documentation) - [Contributing](#contributing) - [Credits](#credits) - [License and legal](#license-and-legal) ## What you need - **Your own *God of War* (USA, `SCUS-97399`) ISO**, about 8 GB. - **~30 GB free** (extracted disc + build). - **64-bit x86 CPU**, 4 cores or more recommended. AVX2 is *not* required. - **OpenGL 4.6** GPU (any recent NVIDIA/AMD, or a recent Intel/AMD iGPU).
| Launcher | `God Of War - PC Launcher_<version>_x64-setup.exe` or the `..._x64-portable.zip`, with the setup wizard | `.deb` / AppImage, with the setup wizard (it drives `install.sh`) | | Status | **Primary target** | **Experimental** | > [!WARNING] > **Linux is experimental.** The Linux path is in the repository (`install.sh`, the GLFW/GL presentation path, > `run_gow.sh`, the `.deb`/AppImage packaging), but it hasn't been validated to the same level: no CI job builds or
[`docs/PLAYING.md`](docs/PLAYING.md). The launcher's tabs and options are in [`docs/LAUNCHER.md`](docs/LAUNCHER.md). ## Performance The game's logic is tied to the PS2's 60 Hz VBlank, so **the game never runs above 60 FPS** and never slows down: late frames are dropped. The window can present at 60, 120 or 144 Hz for lower latency. - **Internal resolution 2× is the sweet spot.** 3× and above are heavy and, on mid-range cards, can stall the presentation. - The CPU side is dominated by the **VU1** (the vector unit). Every optimization pass converts more of it to native
## Why the build is slow and big, and why the PS2 is hard The numbers in [What you need](#what-you-need) (30–90 minutes, ~30 GB) are not padding. They follow from what the installer does and from the machine being ported. **Why the first build takes 30–90 minutes** - The recompiler turns about 6,400 game functions into about 6,400 C++ files, roughly **129 MB of source**. Every MIPS instruction becomes explicit C++ (register access, memory access, branch delay slots, 128-bit multimedia operations),
- jmars125/xcx-recompC · 65 Stars
游戏重编译、游戏游玩
Stars 统计更新于 2026/11/1 12:25:55
查看原始资料
编报时的 AI 中文解读
XCX Recomp 是一款将《异度神剑X 终极版》静态重编译为原生 Windows x86-64 应用的工具，适合拥有合法游戏文件并希望在 PC 上以原生性能游玩的玩家。早期开发阶段，支持新游戏、过场动画和游戏玩法。
编报时的 README 依据
Requirements: Windows 10 or later, an x86-64 CPU with AVX2, a GPU with Vulkan 1.3 and `VK_EXT_external_memory_host`, about 36 GiB free during setup (20 GiB after). 1. Download `XCX-Setup-<version>-windows.zip` from [Releases](../../releases/latest) and unpack it. 2. Start `XCX-Setup.exe`. If SmartScreen warns about an unsigned app, choose *More info*, then *Run anyway*. 3. Pick your game files, the language and an install folder, and press *Start*. The setup builds the game once. 4. Play `xcx.exe` from the install folder.
> **No game files are included.** You need your own legally obtained copy of *Xenoblade Chronicles X: Definitive Edition*, dumped from your own console: the base game and the 2.0.0 update (the add-on content is optional). Instructions for obtaining those files will not be provided. > > The game is built from your files on your machine, so the result contains the game. It is for your own use. Do not share it. ## Status Early development. New Game, cutscenes and gameplay run on Windows with sound, text and the menu, at 60 fps by default (30 to 360 in the settings; above 60 is experimental).
Report problems as issues, following [CONTRIBUTING.md](CONTRIBUTING.md). ## Features - Native code, not emulation - Update 2.0.0 and the Exploration Support Package add-on content - XCX-Setup: builds the game from your own NSP files and `prod.keys` - In-game settings menu (F1) and performance overlay (F10) - Keyboard, mouse and gamepad ## Planned - Linux and the Steam Deck - Mods made for emulators: unpack a mod into a `mods` folder - Custom aspect ratios, including ultrawide - VRR ## Installing
- amirmushichge/pixel-playsHTML · 60 Stars
在线游戏
Stars 统计更新于 2026/11/1 12:25:55
查看原始资料
作者描述
Small browser games as isometric pixel dioramas. No build step, no dependencies.
编报时的 AI 中文解读
这是一个包含多款小型浏览器游戏的合集，每款游戏以等距像素立体模型呈现，使用纯JavaScript编写，无需构建步骤或依赖。适合喜欢休闲小游戏的玩家在线游玩。
编报时的 README 依据
# Pixel Plays Small browser games shown as isometric pixel dioramas. Every game is one page of plain JavaScript with its own software rasteriser: no build step, no dependencies, no sprites or image assets. **Play: https://pixel-plays-ruby.vercel.app/** Concept by [abhi](https://x.com/awwbhi2). Game by [AmirMushich](https://x.com/AmirMushich). Visual reference for the style: shawn.'s [Small plays](https://x.com/zzzzshawn/status/2111546123546661289); no assets or subjects were copied. ## Games | Game | What you do | Input | |------|-------------|-------|
Requires [Node.js](https://nodejs.org/); nothing to install. ```bash node serve.mjs ``` Then open http://127.0.0.1:5300/. Any static file server works, and the site deploys as plain static files. ## Add a game Create `<id>/index.html`, load `../shell.js`, call `PixelPlays.mount({ id, rules })`, set `window.PP_SCENE`, load `../engine.js`, then add the game to `GAMES` in `shell.js`. The page needs a `#mute` button. Put its background track at `<id>/<id>.mp3`. - `shell.js` provides the game menu, the Rules popup, the credits and the background music.
## License The code is under the [MIT License](LICENSE). The music is not: see [MUSIC.md](MUSIC.md).
- wn-mitch/tikz-floraLua · 58 Stars
文档美化、绘画生成
Stars 统计更新于 2026/11/1 12:25:56
查看原始资料
作者描述
Procedural watercolour botanical margin flora for LuaLaTeX
编报时的 AI 中文解读
为LuaLaTeX文档在页边空白处生成植物装饰的宏包，通过Lua和PGF计算绘制水彩风格植物，无外部图片，适用于书籍排版或艺术化文档。
编报时的 README 依据
`MODULES`; the gallery and the tests pick it up, and `tests/test_ornaments.lua` holds it to its box. ```latex \florasetup{layout=split, visitors=0.5, mix={lily=2, umbel=0}, palette={hero=murasaki-shikibu, blooms=warm, leaves=teal}} ``` ### Layouts - `left`: the left margin (the default). - `right`: the right margin. - `centered`: both side margins, with two different plants. - `split`: the outer margin of a two-sided document. It is on the right of odd pages and the left of even ones. - `title`: a crown over a title. Wrap the title in
`\begin{floratitle} ... \end{floratitle}`; without it the first block of lines is crowned. - `surround`: a wreath round the text block, for title pages and contents. The flora fills whatever margins the page has, so give it room with `geometry` (about 5 cm for a side margin, or 3 cm all round for `surround`). It also grows into gaps in the text. A gap between lines taller than about three lines (40 pt), such as a `\vspace` left for a figure, opens across the text's full width and the flora reaches into it. The space below a short last page fills the same way.
\documentclass{article} \usepackage[left=5.5cm,right=2cm]{geometry} % room for the flora \usepackage{tikzflora} \florasetup{seed=7} \begin{document} Text. \end{document} ``` ## Settings `\florasetup{...}` sets keys for the rest of the document, from the page it is on. `\florathispage{...}` sets them for one page only. Mistakes stop the compile with a message naming the key. | Key | Values | Effect | |---|---|---| | `layout` | `left`, `right`, `centered`, `split`, `title`, `surround` | Where the flora grows (below) | | `seed` | a number, or `random` | Which garden; each page varies it |
| `mix` | `{kind=weight, ...}` | How much of each kind: 1 usual, 0 none, 2 about double | | `density` | 0 to 10 | Fuller (above 1) or airier (below 1) | | `visitors` | 0 to 10 | The weight of every visitor at once | | `palette` | `{part=ink, ...}` | What paints the `hero`, `blooms`, `leaves`, `leaflets`, `stems`, `sprays`, `ghost`, `cartouche` | | `spin` | a whole number | Picks other inks for parts set to `random`; the shapes stay | | `ornaments` | `{kind=weight, ...}` | How often each ornament kind is picked for its place; 0 never |
tikz-flora was written with Claude Code, Anthropic's coding agent. The art direction, the tuning against the reference and the choice of what to keep were done by hand. ## Setup You need a TeX distribution with **LuaLaTeX** (TeX Live 2023 or later, MiKTeX, or TinyTeX). Check with `lualatex --version`. The example also uses `lipsum` (`tlmgr install lipsum` on TinyTeX). 1. Download `tikzflora.zip` from the site, or build it with `just dist`. 2. Unzip it. It holds `tikzflora.sty`, `tikzflora-boot.lua`, the generator in `lua/flora/`, and `example.tex`.
- glincker/awesome-selfhosted-aiJavaScript · 53 Stars
AI项目导航
Stars 统计更新于 2026/11/1 12:25:56
查看原始资料
作者描述
Curated list of self-hosted open-source AI: local LLMs, chat UIs, agents, MCP servers, RAG, speech and image generation. Health-checked weekly, dead projects removed automatically.
编报时的 AI 中文解读
这是一个持续更新的自托管开源AI项目精选列表，涵盖本地LLM、聊天界面、智能体、MCP服务器、RAG、语音和图像生成等类别，适合寻找可自托管AI解决方案的开发者或研究者参考。
编报时的 README 依据
# Awesome Self-Hosted AI [](https://awesome.re) > Open-source AI you can run on your own hardware or server. 308 projects across local LLMs, chat interfaces, agents, MCP servers, RAG, speech and image generation. A weekly job checks every project against the GitHub API, and anything archived, deleted or silent for over a year moves to [UNMAINTAINED.md](UNMAINTAINED.md). **308** projects | **307** active in the last 90 days | **10.6M** combined stars | **0** unmaintained | refreshed 2026-10-31
See the [live health table](LIVE.md) for stars, last push and license per project. ## Contents - [Agents and Automation](#agents-and-automation) (27) - [Chat Interfaces](#chat-interfaces) (11) - [Coding Assistants](#coding-assistants) (24) - [Document AI and OCR](#document-ai-and-ocr) (10) - [Fine-Tuning and Training](#finetuning-and-training) (10) - [Home and Personal AI](#home-and-personal-ai) (13) - [Image and Video Generation](#image-and-video-generation) (13) - [LLM Runtimes and Inference](#llm-runtimes-and-inference) (38) - [MCP Servers and Gateways](#mcp-servers-and-gateways) (19)
- wiva-hq/CodeRailTypeScript · 53 Stars
远程开发、模型接口服务
Stars 统计更新于 2026/11/1 12:25:56
查看原始资料
作者描述
Self-hosted remote control for your coding agents: Claude Code and Codex chats, approvals, terminals, files, git and dev-server previews for every computer you own, from the browser or your phone. One Docker container on your own server.
编报时的 AI 中文解读
自托管远程控制工具，用于管理Claude Code和Codex编码代理，支持聊天、审批、终端、文件、Git和浏览器预览，可从浏览器或手机访问。
编报时的 README 依据
<td width="33%" valign="top"><b>🔑 Your agents stay yours</b><br><sub>Each computer runs its own <code>claude</code> and <code>codex</code> with its own login. CodeRail never holds AI keys.</sub></td> <td width="33%" valign="top"><b>👥 Built for teams</b><br><sub>Owner, Operator or Viewer per computer, down to folders and agents. Every change is audited.</sub></td> </tr> <tr> <td valign="top"><b>💬 Agent chats</b><br><sub>Claude Code and Codex over ACP: plans, tool calls, models, modes, background tasks.</sub></td>
ACP adapters, and serves terminals, files, git and the browser. The **server** checks every request against the person's role and keeps the history. ## Get started **1 · Run the server** on any machine with Docker, then put HTTPS in front of port 8080 ([details](https://docs.coderail.codes/self-hosting/deploy)): ```sh git clone https://github.com/wiva-hq/CodeRail && cd CodeRail cp .env.example .env # set SETUP_TOKEN docker compose up -d # the image from ghcr.io; everything it keeps is in ./data ```
The image is [`ghcr.io/wiva-hq/coderail`](https://github.com/wiva-hq/CodeRail/pkgs/container/coderail) (amd64 and arm64). Update with `git pull && docker compose pull && docker compose up -d`. **2 · Create the first admin** at `https://your-site/setup` with that token. **3 · Add a computer.** On a Mac: **Machines → Download for Mac** and approve the code. On Linux: ```sh curl -fsSL https://your-site/install.sh | sh ``` Every build is also on [Releases](https://github.com/wiva-hq/CodeRail/releases), with its checksums.
<a href="https://docs.coderail.codes"><img alt="Docs" src="https://img.shields.io/badge/docs-docs.coderail.codes-18181b?style=flat-square"></a> <a href="https://github.com/wiva-hq/CodeRail/pkgs/container/coderail"><img alt="Server: Docker" src="https://img.shields.io/badge/server-Docker-2496ed?style=flat-square&logo=docker&logoColor=white"></a> <img alt="Computers: macOS and Linux" src="https://img.shields.io/badge/computers-macOS%20·%20Linux-52525b?style=flat-square"> <img alt="Status: early" src="https://img.shields.io/badge/status-0.0.1%20early-f59e0b?style=flat-square">
Everything else (roles, the browser, phones, troubleshooting) is in the **[documentation](https://docs.coderail.codes)**. ## Security, in short Self-hosted, outbound-only computers, no AI credentials, invite-only accounts, permissions checked on the server and again on the computer, signed and notarized Mac builds. Not end-to-end encrypted: your server keeps the history. Read the [security model](https://docs.coderail.codes/reference/security); report vulnerabilities privately per [SECURITY.md](SECURITY.md). ## Development ```sh bun install
- Betelgeuse-AI/BetelgeuseC++ · 50 Stars
模型本地部署
Stars 统计更新于 2026/11/1 12:25:56
查看原始资料
作者描述
A custom fork for frontier-scale mixture-of-experts inference on an RTX 3060 with 12 GB VRAM, 16 GB system RAM and local NVMe.** Betelgeuse keeps dense computation on the GPU and brings routed experts into a bounded VRAM pool only when needed.
编报时的 AI 中文解读
Betelgeuse 是一个针对特定硬件（RTX 3060 12GB VRAM）优化的 llama.cpp 分支，用于在有限显存下运行大规模混合专家（MoE）模型，通过将专家层分页到 NVMe 存储来节省显存。
编报时的 README 依据
Betelgeuse keeps dense computation on the GPU and brings routed experts into a bounded VRAM pool only when needed. This is native C++/CUDA work, not a launcher around an upstream binary. The executables are **`bel-server`** and **`bel-cli`**; the llama C API and internal build target names remain unchanged.
Open `http://127.0.0.1:8080` or use the OpenAI-compatible API. The systemd scope enforces the historical 12 GiB process and zero-swap limits on the 16 GB host; it does not allocate extra memory. Keep the server bound to localhost unless you have separately reviewed authentication and deployment security. For direct interactive CLI chat, use the same memory envelope and paging parameters. No separately started server or `--server-base` is needed: ```bash systemd-run --user --scope --quiet -p MemoryMax=12G -p MemorySwapMax=0 \ build/bin/bel-cli \
not a 3.2 GB dense-only pruning scheme. Original two-shard GGUFs and pack-building indexes are conversion inputs, not extra runtime requirements. ## Build native binaries The paged configuration targets **Linux x86-64, CUDA `sm_86`, RTX 3060 12 GB and local NVMe**. Install a C++17 compiler, CMake, CUDA toolkit/driver, liburing development files and the normal [llama.cpp build dependencies](docs/build.md). Building the embedded terminal logo also requires Python 3, `rsvg-convert` (librsvg) and ImageMagick (`magick` or `convert`); the installed binaries need none of these image tools at
--local-dir /path/to/Alpha-Orionis-XS export BUNDLE=/path/to/Alpha-Orionis-XS ``` The profile seeds likely hot experts; it is neither weights nor a prewarmed cache snapshot. For the tested source files, the dense GGUF is about **3.77 GB** (3.50 GiB of dense tensor payload plus metadata), the per-expert-aligned `experts.bin` layout requires **35,481,714,688 bytes** (35.48 GB), the unchanged PLE companion is **28,800,138,432 bytes** (26.82 GiB), and the profile is **6,920 bytes**. Allow roughly 68 GB of storage for the complete runtime bundle. All non-expert tensors are retained; this is
> **Experimental and partly broken. Prefer `bel-server`.** The paged path is hardware-specific, not a production release or a general promise that any large MoE will fit or run well. CLI behavior, unfamiliar models and configurations outside the documented envelope need independent validation. No MTP configuration is certified for exact-output paging. ## What the numbers mean | Result | Token generation | Evidence and limits | | --- | ---: | --- | | Upstream llama.cpp baseline | **0.9 tok/s** | User-reported on the same hardware; not a new controlled benchmark here. |
- One paged model context per process. Keep `-np 1`; general server multi-user features do not make shared paging state safe for parallel contexts. - The custom CUDA path targets `sm_86`. Other GPUs, CPU-only builds, non-Linux systems and other backends are not certified for paging. - The recorded envelope is `-c 512 -ub 64 -t 6 -ngl 99`, with effective eight-token paging chunks. Larger contexts and different batch sizes need their own memory/correctness checks.