CyberClip / 眼镜蛇：外挂式智能眼镜模块的 Idea 阶段

状态：Idea 阶段 · 日期：2026-04-27 · 组织：Mycelium Protocol / Aura AI
本文性质：开放讨论，欢迎质疑、参与、共建。这不是产品发布，是一个想法的公开记录。

一句话定义

CyberClip（眼镜蛇） = 外挂到任意眼镜/帽子/头带的 AI 传感器模块。靠一根 Cable 连手机供电+传数据，手机是大脑，眼镜只是眼睛和耳朵。

---

§ 1 · Idea 背景

现有 AI 眼镜（Ray-Ban Meta、Frame、OpenGlass）有一个共同矛盾：要塞入电池+芯片+天线，导致重量、发热、续航三角难以同时优化。更深层的矛盾是：大多数人已经有一副自己喜欢的眼镜框，凭什么为了 AI 换掉它？

CyberClip 的答案是：不换。直接夹上去。

把所有算力、网络、AI 模型卸载到已经在你口袋里的手机（或颈挂计算盒），眼镜模块只做最简单的事：采集、按键、传输。

---

§ 2 · 调研精华：开源 AI 眼镜生态现状（2024-2026）

以下调研来自 Gemini Deep Research，提取核心信息如下：

2.1 主要开源方案对比

方案	硬件核心	连接方式	优势	局限
Brilliant Labs Frame	FPGA + Micro OLED	蓝牙 BLE	最接近眼镜形态，SDK 完整（Lua/Python/Flutter）	价格高，闭合生态
BasedHardware OpenGlass	ESP32-S3	蓝牙 BLE	~$25 DIY，代码全开源	蓝牙带宽限制，传视频不可行
LilyGO T-Glass	ESP32-S3 + 棱镜显示	WiFi/BLE	Arduino/MicroPython 支持，现货	有显示屏 → 功耗/重量增加
Mentra OpenSourceSmartGlasses	软件框架	依赖硬件	统一 OS 接口，技能插件化	需要配套硬件

2.2 深圳 ODM 半成品供应链

• 歌尔股份 (Goertek)：高通 AR 平台参考设计，提供镜框模组+底层 SDK，适合企业级合作
• 深圳金卫尔 (GoldenWeald)：Smart AI Glasses PCBA，支持定制，对中小项目友好
• Seeed Studio XIAO 系列：散件模组，适合自定义 PCB，开发者友好度高

2.3 核心技术洞察

OpenGlass vs CyberClip 的关键差异：OpenGlass 用蓝牙传数据，带宽上限约 2Mbps，传 1080P 视频流基本不可行。CyberClip 用 USB OTG（UVC/UAC 协议），即插即用（免驱动），带宽轻松支持 1080P 实时流，彻底解决了蓝牙方案的瓶颈。

---

§ 3 · 方案初稿（CyberClip v0.1 设计规格）

3.1 创始人原始设定（完整记录）

以下是从想法出发时的完整设计约束，一字不改地记录在这里：

1. 适配已有眼镜（不换框）
2. 默认无电池——靠直连 Cable 连接外置电源或直连手机（同时供电+数据，省去 WiFi/蓝牙互联的时间和麻烦）
3. 支持选装小电池（如 300mAh）
4. 默认配置：双侧麦克风、左侧闪光灯、右侧摄像头
5. 无语音唤醒功能，无音乐播放功能
6. 全靠一个键：单击开机、双击拍照、三击录像
7. AI 功能全靠 Cable 直连手机获取服务；没有连接时，就是一个纯拍照/录像的外置眼镜模块
8. 连接后自动：获得拍照图片和视频 → 手机监听语音指令，例如”剪辑刚才的视频，按某思路，发小红书”
9. 可选骨传导模块，挂载到眼镜，通过独立开关控制，用于每日新闻语音播报

终极目标：让普通人可以挂载模块，让自己的眼镜变 AI，自由指定后台模型和 Skill，构造开源生态。（事实上也可以外挂到帽子、头带、头盔等等。）

3.2 硬件架构设计

主控芯片选型

推荐 Realtek RTS5822（或同类 USB 视频控制器），而非 ESP32-S3。理由：不需要做 WiFi/蓝牙，直接将 Sensor 和 Mic 压成 USB 数据流输出，更简单、更省电。

核心模块布局（“积木式”磁吸/卡扣）

左侧镜腿                    右侧镜腿
┌─────────────────────────────────────┐
│  [麦克风L] [闪光灯] ──── [摄像头] [麦克风R] │
│  [Type-C 接口]  ←FPC柔性排线→              │
│  [物理按键]                               │
└─────────────────────────────────────┘
              ↓ USB OTG
           手机 / 颈挂计算盒

• 右侧：800万像素摄像头（Sony IMX219 级别）+ 麦克风
• 左侧：麦克风 + 物理按键（单/双/三击）+ Type-C 接口
• 中间：极细柔性排线（FPC）沿镜架上方走线
• 连接：超软硅胶 Type-C to Type-C 线；推荐配合”颈挂计算盒”而非直插裤兜手机（解决拖拽感）

扩展坞（Pogo Pin 磁吸接口）

• 外挂电池包：磁吸 150~300mAh 模块
• 骨传导包：磁吸震子模块 + 独立物理拨动开关

预估成本

组件	参考成本
USB 视频控制器 IC	¥15-25
800万像素摄像头模组	¥20-35
双麦克风 MEMS	¥8-12
PCB + 柔性排线 FPC	¥15-20
外壳（3D 打印）	¥10-15
Type-C 接口 + 按键 + LED	¥5-8
合计 BOM	¥73-115

3.3 软件交互架构（Phone as the Brain）

眼镜本身是”智障”的，一切 AI 赋予都在手机端开源 App 完成。

脱机模式（仅接充电宝或挂小电池）

主控芯片运行极简逻辑。按键触发写卡指令（需内置 TF 卡槽），完成纯粹的”行车记录仪”功能。

联机模式（连入手机）

眼镜模块 (UVC + UAC)
      │ USB OTG
      ▼
手机 App（开源）
  ├─ 自动挂载检测 → 启动后台服务
  ├─ 数据同步：自动提取 TF 卡新增媒体
  ├─ 音频流处理：眼镜麦克风 → 手机 NPU → 语音指令识别
  └─ Skill 路由网关
       ├─ 内容剪辑 Skill（调用本地模型）
       ├─ 发布 Skill（小红书、微信等）
       └─ 自定义 Skill（JSON 插件格式）

Skill 生态设计：手机 App 支持 JSON 格式 Workflow 导入。高阶玩家配置好”拍照+识别+语音播报”逻辑，打包 JSON 分享到社区，普通人一键导入即可获得同样的 AI 技能。

3.4 工程障碍与应对

障碍	严重程度	应对思路
Cable 拖拽感（头部高频转动）	★★★★ 核心痛点	推荐颈挂计算盒，而非直插裤兜；超软硅胶线材改善手感
麦克风音质（手机麦克风在口袋里）	★★★	眼镜麦克风通过 UAC 协议作为 USB 音频输入，直接绕过手机麦克风
iOS 封闭性（MFi 限制）	★★★	初期专攻 Android / 鸿蒙，iOS 作长期目标
闪光灯眩光（镜片内折射）	★★	改为低亮度红外补光或纯状态指示灯
电池重量配重失衡（300mAh ≈ 6-8g）	★★	使用磁吸模块化设计；左右各挂一半

3.5 开源生态路径（两层）

硬件图纸层（OSHW DPG）

• 开源 3D 打印 STL 文件（适配不同镜架的卡扣、帽子夹、头带固定器）
• PCB 原理图 + BOM 表公开
• 淘宝代工厂可直接接单打样，成本透明

手机端工作流（Software DPG）

• 开源 Android App，支持 JSON Workflow 导入
• 开发者社区分享技能包（Skill Pack）
• 对接 Mycelium Protocol——积分激励贡献者，技能包质量靠社区投票

---

§ 4 · 与 Mycelium Protocol 的关系

CyberClip 的硬件本体是 OSHW（开源硬件），任何人都可以自由制造、改进、销售。Mycelium Protocol 层在其上提供：

• 贡献激励：开发者提交 Skill Pack、上传 STL 改进版，通过积分系统获得社区认可
• 质量过滤：社区投票决定哪些 Skill Pack 进入官方推荐列表
• 数字公共物品（DPG）承诺：核心代码永久开源，不会因为商业化而封闭

---

§ 5 · 现在需要什么

这是 Idea 阶段。以下几件事我还没有答案，希望社区一起来思考：

1. 颈挂计算盒是否有比手机更合适的形态？（树莓派 Zero？RISC-V 小板？）
2. Realtek RTS5822 vs Allwinner V831（带 NPU）——是否值得在眼镜端加一点点本地推理能力？
3. 镜架适配——卡扣 vs 磁吸，哪种对非标准镜架更友好？
4. iOS 路径：有没有人做过 UVC over USB-C on non-jailbroken iPhone 的尝试？
5. 骨传导模块的具体震子选型，有没有体积 ≤ 1cm³ 的推荐？

如果你有想法、有资源、有原型经验，欢迎直接通过博客联系或在社区讨论。

---

§ 6 · 预览图

下图由 Gemini 生成，是对 CyberClip 外观方向的概念展示：

注：这是 AI 生成的概念图，不代表最终硬件形态。

---

这个想法是否值得做？你有什么看法？加入讨论 →

Status: Idea Stage · Date: 2026-04-27 · Organization: Mycelium Protocol / Aura AI
Nature of post: Open discussion. This is not a product launch — it’s a public record of an idea inviting critique, participation, and co-building.

One-Line Definition

CyberClip = A clip-on AI sensor module for any existing glasses / hat / headband. A single USB cable connects to your phone for power + data. The phone is the brain. The glasses are just eyes and ears.

---

§ 1 · Background: Why Does This Exist?

Current AI glasses (Ray-Ban Meta, Frame, OpenGlass) share a common contradiction: to embed battery + chip + antenna, you must compromise on weight, heat, or battery life. The deeper issue: most people already own a pair of glasses they love. Why should they replace them for AI?

CyberClip’s answer: Don’t replace. Just clip on.

Offload all compute, network, and AI models to the phone already in your pocket (or a neck-worn compute dongle). The glasses module does only the simplest things: capture, button, transmit.

---

§ 2 · Research Highlights: Open-Source AI Glasses Ecosystem (2024–2026)

Research via Gemini Deep Research. Key findings extracted below:

2.1 Open-Source Platform Comparison

Project	Core Hardware	Connectivity	Strengths	Limitations
Brilliant Labs Frame	FPGA + Micro OLED	Bluetooth BLE	Closest to normal glasses; full SDK (Lua/Python/Flutter)	Expensive, partially closed
BasedHardware OpenGlass	ESP32-S3	Bluetooth BLE	~$25 DIY, fully open source	Bluetooth bandwidth too low for video
LilyGO T-Glass	ESP32-S3 + prism display	WiFi/BLE	Arduino/MicroPython, in stock	Display adds weight/power
Mentra OpenSourceSmartGlasses	Software framework	Hardware-dependent	Unified OS API, skill plugins	Needs matching hardware

2.2 Shenzhen ODM Supply Chain

• Goertek: Qualcomm AR platform reference design; suitable for enterprise-level partnerships
• GoldenWeald (金卫尔): Smart AI Glasses PCBA; supports customization; accessible to smaller projects
• Seeed Studio XIAO series: Component modules for custom PCB; high developer-friendliness

2.3 Key Technical Insight

OpenGlass uses Bluetooth (~2Mbps max), which makes real-time 1080P video streaming infeasible. CyberClip uses USB OTG with UVC (USB Video Class) + UAC (USB Audio Class) — universally driver-free on Android, and easily capable of 1080P streaming. This single architectural decision eliminates the bandwidth bottleneck that constrains all Bluetooth-based open glasses projects.

---

§ 3 · Draft Specification (CyberClip v0.1)

3.1 Original Design Constraints (Verbatim)

1. Must fit existing glasses (no frame replacement)
2. Default: no battery — powered via Cable direct to phone (simultaneous power + data, eliminates WiFi/Bluetooth pairing friction)
3. Optional: clip-on small battery (e.g. 300mAh)
4. Default hardware: dual microphones (both sides), flash LED (left), 1080P camera (right)
5. No voice wake word, no music playback
6. Single-button interaction: 1-click = power on; 2-click = take photo; 3-click = record video
7. AI features require Cable connection to phone; without connection it’s a pure capture peripheral
8. When connected: auto-sync photos/videos + phone processes audio from glasses mics for voice commands (e.g. “Edit the clip I just shot, summarize it, post to Xiaohongshu”)
9. Optional bone conduction module (clip-on, with dedicated physical toggle switch) for daily news audio broadcast

Ultimate goal: Build an open-source ecosystem where ordinary people can clip on a module, turn their existing glasses AI, and freely specify backend models and Skills. (Works on hats, headbands, helmets too.)

3.2 Hardware Architecture

Controller Selection

Recommend Realtek RTS5822 (or equivalent USB video controller) rather than ESP32-S3. Rationale: no need for WiFi/Bluetooth; directly compresses sensor + mic into a USB data stream. Simpler, lower power.

Module Layout (“Lego-style” magnetic/clip attachment)

Left temple                         Right temple
┌──────────────────────────────────────────┐
│ [Mic-L] [LED Flash] ─FPC─ [Camera] [Mic-R] │
│ [Type-C port]                              │
│ [Action Button]                            │
└──────────────────────────────────────────┘
                  │ USB OTG
                  ▼
         Phone / Neck Compute Dongle

• Right: 8MP camera (Sony IMX219-class) + microphone
• Left: Microphone + action button (1/2/3-click) + Type-C port
• Center: Ultra-thin FPC flexible ribbon cable routed along the top frame
• Cable: Ultra-soft silicone Type-C to Type-C; pair with a neck compute dongle rather than plugging directly into a pocket phone (to reduce cable drag)

Expansion Pogo Pin (Magnetic Attachment)

• Battery pack: Magnetic 150–300mAh module
• Bone conduction pack: Magnetic vibration module + dedicated physical toggle switch

Estimated BOM Cost

Component	Estimated Cost
USB video controller IC	¥15–25
8MP camera module	¥20–35
Dual MEMS microphones	¥8–12
PCB + FPC ribbon cable	¥15–20
3D-printed housing	¥10–15
Type-C + button + LED	¥5–8
Total BOM	¥73–115

3.3 Software Architecture (Phone as the Brain)

The glasses module is “dumb.” All AI capability lives in an open-source phone app.

Offline Mode (Battery pack or power bank only)

Minimal firmware logic: button triggers write-to-TF-card. Pure “action camera” / dashcam mode.

Online Mode (Cable-connected to phone)

Glasses Module (UVC + UAC)
      │ USB OTG
      ▼
Phone App (open source)
  ├─ Auto-mount detection → start background service
  ├─ Media sync: auto-pull new files from TF card
  ├─ Audio stream: glasses mics → phone NPU → voice command recognition
  └─ Skill Router Gateway
       ├─ Clip & Edit Skill (local model)
       ├─ Publish Skill (Xiaohongshu, WeChat, etc.)
       └─ Custom Skill (JSON plugin format)

Skill Ecosystem Design: The app supports JSON-format Workflow import. Power users configure a skill (“capture → identify → voice broadcast”), export as JSON, share to the community. Beginners import and immediately get the same AI capability on their glasses.

3.4 Engineering Obstacles

Obstacle	Severity	Mitigation
Cable drag (head moves constantly)	★★★★ Core issue	Neck compute dongle; ultra-soft silicone cable; route cable along collar
Mic audio quality (phone in pocket)	★★★	Glasses mics act as UAC USB audio input; phone mic bypassed entirely
iOS ecosystem (MFi restrictions)	★★★	Initial focus on Android/HarmonyOS; iOS as long-term target
Flash LED glare (refraction inside lenses)	★★	Replace with low-intensity IR fill light or pure status LED
Battery weight imbalance (300mAh ≈ 6–8g)	★★	Magnetic modular design; split weight across both temples

3.5 Open-Source Ecosystem Layers

Hardware Layer (OSHW DPG)

• Open-source 3D-printable STL files (clips for different frame types, hat clips, headband mounts)
• PCB schematics + BOM published openly
• Taobao/JLCPCB manufacturers can directly produce from files; cost is transparent

Software Layer (Software DPG)

• Open-source Android app with JSON Workflow import
• Developer community shares Skill Packs
• Integrates with Mycelium Protocol: token incentives for contributors, community voting for skill quality

---

§ 4 · Relationship to Mycelium Protocol

CyberClip hardware is OSHW — anyone can freely manufacture, modify, and sell it. The Mycelium Protocol layer adds:

• Contribution incentives: Developers who submit Skill Packs or improved STL files earn community recognition via the points system
• Quality filtering: Community votes determine which Skill Packs enter the official recommended list
• Digital Public Goods (DPG) commitment: Core code stays open-source permanently, no lock-in if commercialized

---

§ 5 · Open Questions (Help Wanted)

This is idea stage. Here are the things I don’t have answers for yet — community input welcome:

1. Neck compute dongle — better form factor than a phone? (Raspberry Pi Zero? RISC-V board?)
2. Realtek RTS5822 vs Allwinner V831 (has onboard NPU) — is any local inference on the glasses worth the complexity?
3. Frame attachment — snap-clip vs magnetic vs adhesive: which is most universal for non-standard frames?
4. iOS path: Has anyone successfully used UVC over USB-C on non-jailbroken iPhone at the app level?
5. Bone conduction module: Any recommendation for a vibration transducer ≤ 1cm³?

If you have ideas, resources, or prototype experience, feel free to reach out or join the discussion.

---

§ 6 · Preview Image

The image below was generated by Gemini as a conceptual visualization of the CyberClip direction:

Note: AI-generated concept art. Does not represent final hardware.

---

Think this is worth building? What’s your take? Join the discussion →