Building a game clipper with AI agents in one day

ClipTokens went from a first cut to on-device voice in a single build day. The versions are 0.1.0 through 0.1.9. There is no 0.1.2. Cloud coding agents wrote the code. GitHub Actions on Windows built the installer and ran smoke tests and screenshots. A person reviewed each release before it went out.

What the day contained

0.1.0 was the MVP: a Windows replay buffer.

0.1.1 fixed Elgato against FFmpeg 8, added a live preview, a 5-minute buffer, and a trim editor.

0.1.3 fixed loopback and Elgato paths, picked an encoder automatically, switched the UI dark, and put the trimmer on a timeline.

0.1.4 added vector icons and thumbnail actions.

0.1.5 stopped the Elgato 4K S from being opened twice. Preview and the buffer share one capture. The preview is 30 fps. Failures show on screen. If NVENC is not there, a software encoder takes over.

0.1.6 starts the buffer on its own, with a watchdog. The wake word stays on. Whisper transcribes. NVENC presets cover 1080p60.

0.1.7 added a microphone picker and a level meter, a chime and an in-game popup when a clip saves, smoother ddagrab display capture, a source-switch fix, and the free watermark.

0.1.8 made the voice talk back, fixed a pause glitch, and put the logo in the watermark.

0.1.9 keeps voice on the PC: Whisper small.en on NVIDIA, titles written on device, and the gold token as the icon.

What you can take today

We share this build log. We do not publish the main source repository.

The open-source fork, which you would build yourself, is not up yet. The Download button on this site is an unsigned free beta. The signed Windows build from main, updated daily, is the supporter plan. Code signing is not set up, so that plan is marked coming soon.