--- language: - zh - en tags: - macos - ai-workbench - local-first - tauri - rust - agent - privacy - llm license: mit pretty_name: 栖木 Perch — Local-first AI Workbench for macOS --- > ⭐ **核心应用** —— 栖木 Perch 的起点。下面是它的完整介绍与下载。 > 开源组件:[推理框架](https://huggingface.co/hongcoding/perch-ultrafast) · [GPU 加速](https://huggingface.co/hongcoding/perch-gpu) · [设计系统](https://huggingface.co/spaces/hongcoding/perch-design) · [完整源码](https://huggingface.co/datasets/hongcoding/perch-source) # 🪵 栖木 Perch ### 你的 AI 助手,凭什么要等云端? > **点发送 → 盯着空白屏幕 → 1 秒、2 秒、3 秒……** > 等它终于开口,你已经分神了。 > > 栖木把这段等待砍到了 **0.08 秒**。 ![栖木 Perch 主界面](./main-ui.webp) **一个真住在你 Mac 里的 AI 工作台。** 不联网也能用,不用把隐私交出去,算力归你。 ⬇️ **[立即下载 v5.2.1](./Perch_5.2.1_aarch64.dmg)**(Apple Silicon · 25.9MB) --- ## ⚡ 三个立刻能感觉到的数字 | | 别人 | 栖木 | |---|---|---| | **首字延迟** | 0.88 ~ 1.39 秒 | **0.08 秒** | | **长文显示** | 写完才一次性蹦出来 | **边写边显示** | | **并发对话** | 只能等一条答完 | **A 在写,B 照常问** | 这不是调参调出来的,是架构差异: > **两条流同时开**——本机 0.6B 小模型先出草稿(0.08s 让你有东西读), > 云端大模型的第一个字一到,立刻接棒替换(质量升级)。 > 你看到的首字 = 本机延迟,最终答案 = 云端质量。**两头都占。** 这个机制叫**混合流**,核心代码已单独开源 → [**perch-ultrafast**](https://huggingface.co/hongcoding/perch-ultrafast)(Rust · MIT) --- ## 🤖 它不只是个聊天框 | 它能干的事 | 说明 | |---|---| | **读写你的文件** | 直接改代码、写文档、整理目录 | | **跑命令** | 执行 shell,18 项危险操作红线 + 全量审计 | | **联网搜索** | 多后端 + 正文摘要 + 提示注入防御 | | **建知识库** | 本地语义检索、引用溯源、专家角色 | | **排班干活** | 计划任务、批量处理、自动巡检 | | **盯着你干活** | 小栖助理只读旁观,随时给建议 | **它是智能体,不是套壳。** --- ## 🔒 隐私不是一句口号(是架构) | 承诺 | 怎么做到 | |---|---| | **数据不上传** | 对话/知识库/记忆默认全在本地 SQLite | | **密钥不落地** | API 凭据存 macOS 钥匙串,绝不写明文 | | **云端只见密文** | 端到端加密同步,主密钥只在你电脑 | | **每一步都留痕** | 工具调用全量审计,敏感信息自动脱敏 | > 你可以一直断网用它。连云都不连,谁也拿不走你的数据。 --- ## 🎯 谁应该试试 - **开发者**:让它读你的仓库、改代码、跑测试 - **写作者**:5000 字长文边生成边读,不用干等 - **隐私敏感者**:本机模型 + 本地存储,数据不出门 - **效率控**:多对话并发 + 快捷键(⌥Space 唤起) --- ## 📦 安装(30 秒) 1. **[下载 `Perch_5.2.1_aarch64.dmg`](./Perch_5.2.1_aarch64.dmg)**(Apple Silicon) 2. 拖进「应用程序」 3. 打开 → 设置 → 填你自己的模型 API > 所有 AI 服务 API **由你自己配置**,不捆绑、不送额度、不抽成。 > 首次打开若提示"未识别开发者" → 右键 → 打开。 支持任意 OpenAI 兼容接口(DeepSeek / 小米 MiMo / 通义 / …)+ 本机 Ollama。 --- ## 🛠️ 技术底子 **Rust 74.7% · TypeScript 18.1% · CSS 6.5%** Tauri 2 + React 19 + rusqlite · **1156+ 测试全绿** · MIT 开源 ## 📸 更多截图 ![专属助理](./assistant.webp) *小栖专属助理:只读工具,盯着会话出主意* ![思考强度滑块](./think-slider.webp) *思考强度滑块:五档连续拖拽* ## 🔗 相关项目 | 项目 | 说明 | 链接 | |---|---|---| | **栖木 Perch**(应用) | 就是你正在看的这个 | https://huggingface.co/datasets/hongcoding/perch | | **perch-ultrafast**(开源框架) | 混合流 + Ultrafast 推理框架(Rust,MIT):首字 1s → 80ms | https://huggingface.co/hongcoding/perch-ultrafast | | **perch-source**(完整源码) | 全部组件开源:133,555 行 / 371 文件 | https://huggingface.co/datasets/hongcoding/perch-source | | **HF Collection** | 上面两个项目归在一起 | https://huggingface.co/collections/hongcoding/perch-ai-workbench-and-ultrafast-framework | ## ⭐ 觉得有意思? - **GitHub 给个 Star**:https://github.com/ghongxv-web/perch - **HF 点个 ❤️**(右上角) - **提需求**:你的需求就是路线图 --- # 🪵 Perch — Your AI, at Your Fingertips (English) ![Perch main UI](./main-ui.webp) **Your AI assistant shouldn't make you wait for the cloud.** > Click send → stare at a blank screen → 1s, 2s, 3s… > Perch cuts that wait to **0.08 seconds**. **A local-first AI workbench that lives on your Mac.** Works offline. Keeps your data yours. - ⚡ **0.08s first token** — local 0.6B answers instantly, cloud model seamlessly takes over - 🔒 **Privacy by architecture** — everything local by default; optional E2E-encrypted sync - 🤖 **A real agent** — reads/writes files, runs commands, searches the web, schedules tasks - 🧵 **Concurrent conversations** — talk to A while B is still generating - 🧠 **Bring your own models** — local Ollama + any OpenAI-compatible API **Rust 74.7% · Tauri 2 + React 19 · 1,156+ tests · MIT** ⬇️ **[Download Perch_5.2.1_aarch64.dmg](./Perch_5.2.1_aarch64.dmg)** (Apple Silicon) All AI APIs are self-configured. Star on GitHub: https://github.com/ghongxv-web/perch --- **栖木 Perch** — 让 AI 栖息于你的本机,安静而强大。