Skip to content

Install Windows

github-actions[bot] edited this page Nov 21, 2026 · 5 revisions
  1. Tools

    winget install astral-sh.uv
    winget install Gyan.FFmpeg      # the "full" build: libass captions, loudnorm, QSV/NVENC

    Open a new terminal so both are on PATH.

  2. Code: copy the repo folder (everything except streams/, .venv/, logs/) to C:\clipper (or anywhere), then in it:

    uv sync --extra gpu --extra faces --extra discord   # worker + serve + bot; a bot-only machine needs only --extra discord

    The gpu extra pulls torch 2.8 with CUDA 12.8 from the PyTorch index (pinned in pyproject.toml). The CUDA toolkit is not needed; the wheels bring the runtime and cuDNN. If clippa doctor says torch sees no CUDA GPU, the CPU wheel slipped in:

    uv pip install --reinstall torch==2.8.0 torchaudio==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128
  3. Config: copy clipper.toml from the main machine. Things to change per machine:

    • [paths] streams: the NAS share, e.g. "//nas/clipper/streams" or a mapped drive. Every machine must see the same streams folder.
    • [cut] clip_codec: h264_qsv needs an Intel GPU; on an NVIDIA-only desktop use h264_nvenc (driver 610+ for ffmpeg 8.1) or libx264. clippa doctor test-encodes it.
    • [transcribe] compute_type = "int8_float16" and batch_size = 4 on cards under 8 GB. Anything can also be set per machine in clipper.local.toml (what clippa setup writes) or with an env var CLIPPER_<SECTION>_<KEY>, e.g. CLIPPER_CUT_CLIP_CODEC=libx264, so one shared clipper.toml can stay identical.
  4. Secrets: copy secrets/ (discord_token.txt, youtube_client_secret.json, youtube_token.json; only the machine that uploads needs the YouTube files).

  5. Keys as user environment variables (System > About > Advanced system settings > Environment Variables, or):

    [Environment]::SetEnvironmentVariable('ANTHROPIC_API_KEY', '<key>', 'User')
    [Environment]::SetEnvironmentVariable('KITSU_TOKEN', '<Kitsu bot token>', 'User')

    Add ANTHROPIC_WORKSPACE_ID if the key isn't scoped to a workspace. Open a new terminal.

  6. Fonts: thumbnails and Short captions use Impact and Bahnschrift, both standard on Windows. On Linux install ttf-mscorefonts-installer (Impact) and Roboto Condensed (fonts-roboto-unhinted). The weekly issue needs a few more.

    Optional: cover layouts and photos, the cast library, faces.

  7. Check: clippa doctor --role gpu (or --role serve, --role bot).

  8. Run unattended. Serve and the bot run as containers (Docker compose, deploy/hub, on the NAS or any Docker host: Roles and services). On Windows the scheduled task is for the GPU worker (at logon, restarted if it exits, no admin needed; the scripts share deploy\common.ps1, which lists the user env vars the task copies in, since scheduled tasks don't always see ones set after logon: a new key goes there too):

    deploy\worker\worker-task.ps1 -Install     # GPU worker

    -Status, -Restart, -Stop, -Uninstall work. Logs go to logs\. clippa services shows the task plus the recorder and any NAS cuts, with serve's and the bot's heartbeats from the share; clippa services restart worker runs the same script, and serve or bot over ssh once [serve] hub names the compose folder. The bot reaches serve at [serve] url (empty = this machine); on different machines set it, give both the same [warroom] token (or secrets/warroom_token.txt), and put [paths] schedule and [paths] news under the streams share on both (the bot reads what serve writes there; clippa doctor --role bot checks).

Clone this wiki locally