tutor_vision

A one-key HDMI screenshot tutor for Linux. Developed on a Dell E6400 (Core 2 Duo, 64-bit Linux), but works on any Linux machine with a V4L2 MJPEG capture device.

Run the agent, point an HDMI source at the USB capture dongle, press Space, and it grabs a fresh MJPEG frame, sends it (base64 as a data URL) to a vision model via OpenRouter, and prints the tutor's answer. Press q or Ctrl-C to quit.

What you need

• Any USB HDMI capture dongle that exposes a V4L2 MJPEG device (e.g. MS2109).
• Build only: a POSIX box with gcc, make-style tooling, and static libcurl/libssl/libcrypto/libz archives.
• An OpenRouter API key and a vision-capable model.

Setup

Copy the key template and put in your real OpenRouter key (this file is gitignored):

cp secret.h.example secret.h
# edit secret.h

Build

./build.sh

Produces a single static agent binary whose only dynamic dependency is libc (carries its own TLS/HTTPS via static curl + OpenSSL). Copy it to your target machine and run:

./agent

On startup it opens /dev/video0, sets 1280x720 MJPEG, and streams so the HDMI source stays connected.

Configuration

• Model: set TUTOR_VISION_MODEL env var to override the default (google/gemini-2.5-flash).
• Video device: edit VIDEO_DEV in agent.c if your capture device lands on a different /dev/videoN.