Show HN: Claude Code built my ESP32 and fixed its UI by looking at a webcam

**Show HN: Claude Code built an ESP32 project and fixed its UI from a webcam feed**

A Show HN submission titled “Claude Code built my ESP32 and fixed its UI by looking at a webcam” points to the GitHub repo **fcavalcantirj/claude-code-eyes**, with discussion at news.ycombinator.com/item?id=48997881. At capture it had **1 point** and **0 comments**. The claim is concrete: **Claude Code** was used not only to build firmware or tooling for an **ESP32**, but to correct the device’s **UI** by inspecting a **webcam** view of the running system rather than relying only on logs or static screenshots.

The technical angle is a closed visual loop. Instead of stopping at compile errors or serial output, the agent is given a live or near-live image of the board’s display (or host UI driving it) and uses that pixels-level signal to decide what is wrong and what to change. That sits next to more common agent workflows—edit code, run tests, read stderr—and adds **computer vision** as a feedback channel for hardware and embedded UI work, where “does it look right on the glass” is often the real acceptance test and is hard to encode as a unit assertion.

For engineers and builders, the useful part is the failure mode this targets. ESP32 projects routinely mix firmware, display drivers, layout, and timing; a misaligned widget, wrong color, clipped text, or dead control may never fail a test suite. Letting an agent **see** the device shortens the “flash, look, guess, patch” cycle that still eats evenings on desk hardware. It also forces a clearer contract: the model needs a stable camera frame, enough resolution to read the UI, and a way to map visual defects back to files and symbols in the repo.

In market terms this lands in the same HN lane as other coding agents (**Claude**, **Codex**, and adjacent tools often grouped with demos like **Fable**): not a new silicon vendor story, but a demonstration of how far agent loops stretch when the environment is physical. Most agent marketing still centers on pure software repos. A public ESP32 + webcam UI fix is a small but sharp proof point that hardware-adjacent work is in scope if sensors (here, a camera) close the loop. Early traction on the thread is minimal (**1** point, no comments yet), so this is still a signal from the author and repo, not a crowded consensus thread.

Practical takeaway: if you ship embedded or kiosk UIs, treat **visual regression from a real camera** as a first-class agent input alongside build logs—mount a fixed webcam, define what “good” looks like in a few reference frames, and keep the agent’s edits tied to the same repo that flashes the board. What to watch next on **claude-code-eyes** is whether the author documents the exact vision path (how frames are captured and fed to Claude Code), how much of the ESP32 stack was generated versus guided, and whether others reproduce the same “see UI → patch → reflash” loop without hand-holding. Until that is spelled out in the repo, treat the Show HN as a directed case study, not a turnkey recipe.