The open‑source skill Claude Code Eyes now gives Claude Code direct visual feedback from a physical screen or hardware setup using any snapshot camera – for example, an old Android phone running IP Webcam – and feeds the image back into the AI for automated verification tweet.
Visual feedback loop
Claude Code captures a frame, scans it for visual defects such as missing characters, clipped text, or mis‑wired pins, and then the agent can edit the source, reflash the firmware, and take another picture to confirm the fix. The workflow replaces manual screenshot uploads with a closed‑loop process that runs until the camera reports a clear, stable image.
Practical impact for startups
For teams building embedded devices, the biggest pain point is the manual step of confirming that a change actually manifested on the hardware. By automating that check, developers can run faster iteration cycles and catch regressions that unit tests miss. The skill works with any cheap camera, so there’s no need for specialized vision hardware; the main requirement is a reliable network connection to stream frames.
Limitations and cost considerations
The system depends on image quality – blurry or poorly lit frames are automatically rejected, which can add retries and increase overall test time. It also only detects visual anomalies; electrical faults that don’t show up on the display remain invisible. Since the skill is open source, there are no licensing fees, but you’ll need a device to run the camera and enough compute to run Claude Code’s inference, which may be a factor on constrained CI runners.
When to try it
If your startup ships firmware that renders UI or status LEDs, set up a cheap Android phone as a camera and run Claude Code Eyes on a staging branch. Watch for false positives from lighting changes and measure the net time saved versus manual verification before rolling it into your main CI pipeline.