VisionClaw
VisionClaw is a public-source prototype that turns Meta Ray-Ban smart glasses into a voice-and-vision agent surface: glasses stream visual/audio context to Gemini Live, then optional tool calls route through an OpenClaw gateway. Source: Intent-Lab/VisionClaw README, accessed 2026-07-04
Source snapshot
Intent-Lab/VisionClaw describes itself as a real-time assistant for Meta Ray-Ban smart glasses on iOS and Android. The pipeline streams glasses or phone-camera frames at roughly 1fps plus bidirectional audio into Gemini Live; Gemini can then call a single execute tool that the app forwards to OpenClaw for connected-app actions such as messaging, search, notes, reminders, and smart-home tasks. The repo also includes phone-camera fallback mode and a WebRTC viewer for sharing the glasses point of view. Source: Intent-Lab/VisionClaw README, 2026-07-04
GitHub snapshot on 2026-07-04: Intent-Lab/VisionClaw, default branch main, 2,430 stars, repo created 2026-02-06, latest commit 1d6ca07 on 2026-05-06, and no tagged releases returned by gh release list. GitHub reports the license as NOASSERTION; the LICENSE file points users to Meta Wearables Developer Terms and Acceptable Use Policy. Treat this as public research/prototype code with terms constraints, not a broadly permissive open-source dependency.
Why it matters
VisionClaw is a concrete wearable-agent architecture rather than another chat wrapper:
- the interface is glasses plus voice, with the phone app as control plane;
- the model loop is Gemini Live over WebSocket, not speech-to-text bolted onto a text chat;
- the action layer is optional OpenClaw, so tool execution can reuse a local gateway and connected personal apps;
- the test mode uses the phone camera, which lowers hardware friction for development;
- the product risk is privacy, consent, and action authorization because camera, microphone, location, and messaging can all meet in one loop.
The durable pattern belongs beside Computer Use and Browser Automation Patterns and Agent Product Surface: agents are moving from browser/desktop automation into first-person, always-nearby sensor surfaces. The same review controls still apply. A wearable agent needs visible session state, hard stop/cancel behavior, permission scoping, and proof of side effects before it touches messaging or home automation.
Caveats
VisionClaw is useful as an architecture signal, not as a default install. It depends on Meta Wearables DAT SDK access, a Gemini API key, platform-specific mobile builds, and optional OpenClaw gateway configuration on the same network. Android setup also requires GitHub Packages access for the DAT SDK. The README's quick-start clone command still references sseanliu/VisionClaw while the canonical reviewed repo is Intent-Lab/VisionClaw; use the latter unless the project moves again.
Timeline
- 2026-07-04 | Created from the X bookmark artifact review and primary-source GitHub inspection. Captured the Gemini Live + Meta DAT + optional OpenClaw architecture, phone fallback, WebRTC viewer, non-standard license status, and hardware/privacy caveats. Source: X/@ihtesham2005, 2026-03-28; Source: Intent-Lab/VisionClaw