ClipTokens
Free beta Download for Windows. Unsigned for now. Opens the GitHub release in a new tab.

Built in public

How ClipTokens is built.

This is the open blueprint of an AI-engineered Windows app, quirks included. We share how it is built. We do not publish the main source repository.

Daily notes land on main first. Supporters, when the signed build exists, get that main branch as a Windows app updated every day. Stable releases follow. The installer on this site today is an unsigned free beta. The open-source fork, which you build yourself, is not published yet.

Architecture

The app
.NET 8 WPF. One Windows desktop app.
FFmpeg
FFmpeg 8.1.3, LGPL, runs as its own process. It is not linked into the app.
One capture
A single capture process feeds both the replay buffer and a 30 fps preview.
Desktop capture
ddagrab, Desktop Duplication, with zero-copy into NVENC. If that path is unavailable, gdigrab is the fallback.
Encode
NVENC, with HEVC and AV1 presets, and a software encoder fallback.
Save
A buffer state machine with a watchdog. A stream-copy remux writes the clip in about 0.1 seconds.

On-device voice

Free voice runs entirely on your PC. This website does not receive it.

  • The wake word is local: “hey clip”.
  • Whisper small.en runs on NVIDIA through CUDA. base.en runs on CPU.
  • The model stays warm.
  • Endpointing is fast.
  • Windows TTS talks back, and barge-in settles the title.

How it is engineered

  • Cloud coding agents write the code.
  • GitHub Actions on Windows builds the installer and runs smoke tests and screenshots.
  • A human reviews before each release.

Connecting an advanced model, for titles, a smarter conversational assistant, and auto upload, is the supporter plan. It needs an account, and it is not for sale yet.

Credits

Backer names and founding partner logos will be listed here after someone actually backs the blueprint. None yet.

Quirks

Real bugs from the 0.1 run, and what fixed them.

  1. The Elgato 4K S was opened twice, once by preview and once by the buffer, and the device came back busy.

    One shared capture process.

  2. FFmpeg 8 rejected an old option.

    The option was dropped for the FFmpeg 8 command.

  3. The buffer failed to start and stayed quiet. Five startup races were involved.

    The state machine and watchdog now surface the failure instead of swallowing it.

  4. Switching source left the new session paused. An old process exiting marked the new session dead.

    A session only dies when its own process exits.

  5. Windows dictation was almost always wrong.

    Transcription moved to Whisper on the PC.

  6. The wake word missed quiet microphones.

    A mic picker and gain.

  7. The CI runner has no GPU.

    NVENC, ddagrab, and Elgato are tested on a real PC. CI still builds the installer and runs the checks it can.

  8. The installer is unsigned, so Windows SmartScreen warns.

    Still true for the free beta. Code signing is not set up, which is why the signed supporter build is marked coming soon.