桌面养只老鼠 · AI 帮你干活 A pixel-art desktop AI that lives in your screen corner
v0.4 · 12 MB 安装包 · 首启下载 ~260 MB 语音模型 · 100% 本地推理 · 国内无需 VPN
一只睡在你 macOS 屏幕角落的像素老鼠。
- 按住 ⌘⇧Space + 说话 → 它截屏 + 听你说 + 调本机的 AI（Claude/Codex/OpenClaw/Hermes）→ 流式回答到气泡
- 长按 fn + 说话 → 中英文语音直接打到任何输入框的光标位置（不切输入法）
- 拖文件给它 → 老鼠"吞下"后你接着说"用三句话总结" → AI 读懂文档帮你
100% 本地推理（sherpa-onnx 语音识别 + 标点模型，零字节上云），但 AI 调用部分用你自己装的 CLI（你的 token 你做主）。
30-秒功能演示 · ▶ 看完整版 MP4
- 零切换：没窗口要打开，没 app 要 focus，老鼠就在那
- AI 后端你选：装了 Claude / Codex / Hermes / OpenClaw 都行，托盘菜单切换
- 圈定问题：按住快捷键 + 按左键拖动画圈 → AI 拿到的截图带你高亮的部分。试试在 stack trace 上圈
- 认识你的项目：托盘指定 workspace path 一次 → AI 的 Read/Write/Bash 都以这个为 cwd。"修这个 bug" 真能修
- 能动手：说"打开附近的咖啡馆" → 开 Maps。"把这个网页转成 Word" → 干。macOS URL schemes + osascript + Claude Bash 工具 = AI 真的能做事
- 浏览器自动化：托盘一键开
chrome-devtools-mcp，AI 能驱动你的真实 Chrome（独立调试 profile） - 历史会话：托盘"📜 查看历史" → 点任意旧对话「💬 继续这个话题」→ 接着聊不丢上下文
- 📦 DMG 12 MB（之前 238 MB）—— 模型首启动按需下载，国内优先 hf-mirror.com / gh-proxy.com
- 🌐 中文 + 英文双模型：Onboarding 选语言，海外用户用英文模型（73 MB），中文用户用 zh-en bilingual（199 MB）
- 📥 多镜像 fallback + 断点续传：网络抖动自动切镜像 / 断网重启接着下 / 30s 速度<1KB/s 主动断开
- 🐭 陪伴向动画全集：眼球追鼠标 · 打字时点头陪伴 · 贴近抬头 · 闲置渐睡 · 摸头浮爱心 · 亲密度成长（互动越多眼睛越大）· 伸懒腰/晕/小跳/深夜困/冷落委屈 —— 9 款皮肤全适配
- 🧱 撞墙回弹：桌宠跟随光标 / 被拖动 / 气泡撞到屏幕边时被挡在屏内 + 软果冻挤压回弹，不再飘出屏幕或乱飘
- 📋 Reactive 反应：复制内容 → 桌宠抖耳 + 头顶弹一排快捷动作（翻译 / 解释 / 📄 纯文本…），5 秒自动消失或点了就做
- 🍽 拖文件喂桌宠：Finder 文件拖到老鼠嘴里 → 它"吞下" → 你接着说"总结一下" = AI 读 PDF/网页/截图/代码
- 🧠 AI 实时活动：AI 处理期间气泡显示"正在思考 / 读文件 / 跑命令"，不再像卡死
- 🔁 AI 任务串行队列：一次做一件、排队逐个做完都通知；语音听写永远即时不排队；忙碌时桌宠右下角三点 badge
- 🎓 首次使用引导：模型下完桌宠主动跳出教学（含 learn-by-doing 真用一遍）· PetMenu「📖 教我用」可重看
- ✏️ 语音确认可编辑：转写完进确认框可直接改字 · Esc 取消 / Enter 立即 — 防止语音误识别浪费 token
- 📖 语音术语库 + 本地纠错：内置程序员词表（API / Tauri / Rust / commit 等不被听错）+ 可加自定义词；3 秒规则纠常见误识别，纯本地不调 LLM
- 🔤 选词触发：选中文字也能弹 Reactive 动作（macOS 限制：仅原生 app，Chrome/Electron 用复制路径）
- 🦞 桌宠跟随光标：托盘开关，老鼠在屏幕里跟着你的鼠标走
- 🎨 9 款皮肤：经典灰 / 小白鼠 / 田鼠 / 忍者 / 机械 / 金鼠 / 🐱 小灰猫 / 🦊 赤狐 / 🐸 树蛙
- 语音识别 100% 本地：sherpa-onnx Zipformer 模型在你电脑上跑，不联网
- 剪贴板 AES-256-GCM 加密，密钥存 macOS Keychain（只有你能读）
- 标记为 transient 的内容直接忽略：1Password / Bitwarden / sudo 提示 → 永远不记
- 截图保存在临时文件，可以
open查看 - AI 调用走你自己的 CLI：你装了 Claude Code 就用你的 Anthropic key；装 Codex 就用 OpenAI；不上 MouseClaw 的云
最简单：下最新 .dmg → 拖到 Applications。
⚠️ Developer-ID 签名但未公证。首次启动：右键 MouseClaw.app → "打开" → Gatekeeper 警告 → 仍打开。
首启 onboarding：
- 选快捷键（默认
⌘⇧Space） - 选 AI 后端（缺哪个 onboarding 会告诉你怎么装）
- 选桌宠皮肤
- 选语音模型语言 + 语音输入触发键
- 桌宠位置 + 授权（辅助功能 / 屏幕录制 / 麦克风）
完事后桌宠开始下载语音模型（zh-en ~199MB / en ~73MB + 标点 ~62MB），下完会主动教你用一次。
macOS 11+ · Apple Silicon 推荐（Intel 没仔细测）· Windows/Linux 暂无
Pre-1.0，迭代快。看 CHANGELOG 知道最新。
A pixel-art mouse that sleeps in the corner of your macOS screen.
- Hold ⌘⇧Space + speak → screenshots + transcribes + asks your AI CLI (Claude / Codex / OpenClaw / Hermes) → streams reply into a speech bubble
- Hold fn + speak → voice transcription typed directly into your cursor in any app (no IME switching, English & Chinese)
- Drag a file onto the pet → it "eats" the file, then you say "summarize in 3 sentences" → AI reads it
100% local speech recognition (sherpa-onnx + punctuation model, zero bytes uploaded). AI calls use your own CLI (your tokens, your rules).
30-second tour · ▶ Watch full MP4
- Zero context switch. No window to open, no app to focus
- Pick your AI brain. Claude / Codex / OpenClaw / Hermes — switch in tray menu
- Circle the bug. Hold shortcut + left-click + drag to highlight on screen. AI gets screenshot with your marks
- Workspace-aware. Tray sets project folder → AI's Read/Write/Bash use it as cwd. "Fix this bug" actually fixes it
- Computer use built-in. macOS URL schemes + osascript + Claude's Bash tool → AI actually does things
- Browser automation. Tray enables
chrome-devtools-mcpagainst dedicated profile of your real Chrome - Session history. Tray → "📜 View history" → "💬 Continue this conversation" picks up where you left off
- 📦 12 MB DMG (was 238 MB) — models download on first launch with multi-mirror fallback
- 🌐 Chinese + English — pick at Onboarding; English-only users get 73 MB model (vs 199 MB bilingual)
- 📥 Resumable downloads — auto-switch mirror on failure, resume from
.parton crash/restart - 🐭 Full companion animations — eyes track the cursor · nods while you type · perks up on hover · dozes off when idle · heart on pat · intimacy growth (bigger eyes the more you interact) · stretch/dizzy/hop/late-night-drowsy/neglected sulk — across all 9 skins
- 🧱 Edge bounce — following the cursor / dragged / bubble hitting a screen edge: the pet stays on-screen with a soft squash-and-recoil, no more drifting off
- 📋 Reactive ribbon — copy something → the pet twitches + a row of quick actions pops above its head (translate / explain / 📄 plain text…), auto-dismiss in 5s
- 🍽 Feed files — drag a file onto the pet, it "eats" it, then ask "summarize this" — AI reads PDF/page/screenshot/code
- 🧠 Live AI activity — bubble shows "thinking / reading file / running command" during long tasks, never looks frozen
- 🔁 Serial AI queue — one task at a time, queued & each notified on done; voice typing never queues; busy badge on the pet
- 🎓 First-run tour — pet teaches you after model downloads (with hands-on learn-by-doing); "📖 Teach me" in pet menu to redo
- ✏️ Editable voice confirm — edit the transcript in place after transcription, Esc cancel / Enter send (prevents token waste on misrecognized speech)
- 📖 Voice vocabulary + local correction — built-in programmer terms (API / Tauri / Rust / commit…) + your own custom words so jargon isn't misheard; rule-based correction of common misrecognitions, 100% local, no LLM
- 🔤 Selection trigger — select text to pop the Reactive actions too (macOS limit: native apps only; Chrome/Electron use the copy path)
- 🦞 Cursor follow — pet follows your mouse around screen (toggle in tray)
- 🎨 9 skins — classic / white / brown / ninja / robot / gold / 🐱 cat / 🦊 fox / 🐸 frog
- Voice recognition 100% local. sherpa-onnx Zipformer runs on your CPU/Metal
- Clipboard AES-256-GCM encrypted with key in macOS Keychain
- Transient-marked content ignored (1Password, Bitwarden, sudo prompts)
- Screenshots in tmpfs,
opento inspect - AI calls use your CLI — your Anthropic / OpenAI key, never through us
Download latest .dmg → drag to Applications.
⚠️ Developer-ID signed but not notarized yet. First launch: right-click → "Open" → Gatekeeper → Open.
First-launch onboarding picks shortcut, AI backend, skin, voice language, voice-IME trigger, pet anchor, permissions. Then downloads voice models (~199 MB zh-en / ~73 MB en + 62 MB punctuation).
macOS 11+ · Apple Silicon recommended (Intel less tested) · Windows/Linux not yet.
Pre-1.0, moving fast. See CHANGELOG.
- Tauri 2 (Rust + React/TypeScript)
- sherpa-onnx streaming Zipformer for ASR (CPU + CoreML)
- sherpa-onnx CT-Transformer for Chinese punctuation
- CGEventTap for
fnlong-press detection - NSPasteboard polling @ 500ms (Maccy's pattern)
- AES-256-GCM + Keychain for clipboard at rest
- chrome-devtools-mcp bridge for browser automation
- Curl with
-C -for resumable model downloads + multi-mirror fallback - Multi-backend: Claude Code CLI / OpenAI Codex CLI / OpenClaw / Nous Hermes Agent
App: 12 MB DMG · 35 MB installed · ~260 MB voice models (downloaded first launch) · ~400 MB RAM idle (incl. loaded models) · <1% CPU idle.
Bug reports + screenshots welcome at Issues.
This is a personal "vibe project" — I make it pretty first, useful second. PRs that match the design direction (see DESIGN.md) are merged fast.
MIT