Voxpip
EzekielOgunkunle/voxpip
Transcreva vídeos localmente — sem API key, sem custos
Ver no GitHubPlugin Claude Code que transcreve vídeos localmente sem necessidade de API keys. Suporta transcrição via STT local (voxtype, whisper.cpp, mlx_whisper), com fallback para APIs em nuvem (Groq/OpenAI). Extrai frames inteligentemente e passa tudo ao Claude para análise contextualizada de conteúdo visual e áudio.
Instalação
npx convez-ai EzekielOgunkunle/voxpip1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise de conteúdo de vídeos do YouTube, Loom ou redes sociais
- Extração de transcrições e insights de apresentações ou webinars
- Processamento offline de vídeos sensíveis ou privados
Veredito

Solução prática que resolve um gap real: transcrição de vídeos sem custos de API. Arquitetura clara com fallbacks inteligentes (local STT → Groq → OpenAI). Recomendado para quem trabalha muito com vídeo e quer manter dados offline.
README
A tradução para o português deste README ainda está na fila.
/voxpip
Give Claude the ability to watch any video — even without an API key.
Claude Code:
/plugin marketplace add EzekielOgunkunle/voxpip
/plugin install voxpip@voxpip
claude.ai (web): install via the Claude Code plugin marketplace or clone the repo directly.
Codex / generic skills:
git clone https://github.com/EzekielOgunkunle/voxpip.git ~/.codex/skills/voxpip
Zero config to start — local STT detected automatically on first run.
Note: voxpip is a Claude Code plugin, not a PyPI package.
pip install voxpipwill not work — use the install commands above.
I built this because I work with a lot of video content and I wanted a way to hand video to Claude without needing an API key just to get a transcript. Brad Bonanno's /watch plugin was the perfect foundation — solid pipeline, clean code. I added a local STT tier so the transcription fallback works even if you've never touched a Groq or OpenAI account.
What it adds to /watch
This plugin forks bradautomates/claude-video. Brad's pipeline handles everything: download via yt-dlp, frame extraction via ffmpeg, native caption parsing, and cloud Whisper fallback (Groq/OpenAI). voxpip's single addition is a local STT tier that sits between caption parsing and the cloud APIs.
If a video has no native captions, voxpip tries your local STT engine before reaching for a cloud API. If local STT isn't installed either, it falls back to Groq or OpenAI exactly as the upstream does. You can also force the behavior with --no-local-stt or --no-whisper.
How it works
- You paste a video and a question. URL (anything yt-dlp supports — YouTube, Loom, TikTok, X, Instagram, plus hundreds more) or a local path (
.mp4,.mov,.mkv,.webm). yt-dlpdownloads it. For URLs, into a temp working directory. For local files, no download — just probed in place.ffmpegextracts frames at an auto-scaled rate. Duration-aware budget: ≤30s gets ~30 frames, 30-60s gets ~40, 1-3min gets ~60, 3-10min gets ~80, longer gets 100 sparsely. Hard caps: 2 fps, 100 frames. JPEGs at 512px wide by default.- The transcript comes from one of three places. First: native captions from yt-dlp (free, instant). Second: local STT engine on your machine (free, offline). Third: Whisper API — Groq's
whisper-large-v3(preferred) or OpenAI'swhisper-1. - Frames + transcript are handed to Claude. The script prints frame paths with
t=MM:SSmarkers and the transcript with timestamps. ClaudeReads each frame in parallel — JPEGs render directly as images in its context. - Claude answers grounded in what's actually on screen and in the audio. Not "based on the description" or "according to the title." It saw the frames. It heard the transcript.
- Cleanup. The script prints a working directory at the end. If you're not asking follow-ups, Claude removes it.
Local STT engines
| Engine | Platform | Install |
|---|---|---|
| voxtype | Linux (Arch) | paru -S voxtype or yay -S voxtype — non-Arch: see voxtype.io |
| whisper.cpp | Linux/macOS/Windows | github.com/ggerganov/whisper.cpp — build or grab a release binary |
| mlx_whisper | macOS (Apple Silicon) | pip install mlx-whisper |
| whisper (openai-whisper) | Linux/macOS/Windows | pip install openai-whisper |
voxtype note: voxtype is primarily a push-to-talk dictation tool. voxpip invokes it via
voxtype transcribe <file>— voxtype must expose this CLI transcription mode to be usable here. See voxtype.io/docs to verify your version supports it.
setup.py detects whichever is on your PATH and caches the result in ~/.config/voxpip/.env (key: LOCAL_STT_ENGINE). Re-detection only happens if the key is missing.
Bring your own keys (optional)
| Capability | What you need | Cost |
|---|---|---|
| Local STT (auto-detected) | voxtype / whisper.cpp / mlx_whisper / openai-whisper | none required |
| Download + native captions | yt-dlp + ffmpeg | Free |
| Whisper fallback (preferred) | Groq API key — whisper-large-v3 | Cheap, fast |
| Whisper fallback (alt) | OpenAI API key — whisper-1 | Standard pricing |
| Disable Whisper entirely | --no-whisper | Free, frames-only when no captions |
Install
| Surface | Install |
|---|---|
| Claude Code | /plugin marketplace add EzekielOgunkunle/voxpip then /plugin install voxpip@voxpip |
| claude.ai (web) | Install via the Claude Code plugin marketplace |
| Codex | git clone https://github.com/EzekielOgunkunle/voxpip.git ~/.codex/skills/voxpip |
| Manual / dev | git clone https://github.com/EzekielOgunkunle/voxpip.git ~/.claude/skills/voxpip |
Usage
/voxpip https://youtu.be/dQw4w9WgXcQ what happens at the 30 second mark?
/voxpip https://www.tiktok.com/@user/video/123 summarize this
/voxpip ~/Movies/screen-recording.mp4 when does the UI break?
/voxpip https://vimeo.com/123 what tools does she mention?
Focused on a specific section — denser frame budget, lower token cost:
/voxpip https://youtu.be/abc --start 2:15 --end 2:45
/voxpip video.mp4 --start 50 --end 60
/voxpip "$URL" --start 1:12:00 # from 1h12m to end
Other flags:
--no-local-stt— skip local STT tier, go straight to API fallback.--max-frames N— lower the frame cap for a tighter token budget.--resolution W— bump frame width to 1024 px when Claude needs to read on-screen text.--fps F— override the auto-fps calculation (still capped at 2 fps).--whisper groq|openai— force a specific Whisper backend.--no-whisper— disable cloud transcription entirely.--out-dir DIR— keep working files somewhere specific.
Structure
.
├── SKILL.md # skill contract — loaded by all three surfaces [changed]
├── commands/
│ └── voxpip.md # slash command shim [renamed from watch.md]
├── scripts/
│ ├── watch.py # entry point — orchestrates download → frames → transcript
│ ├── download.py # yt-dlp wrapper
│ ├── frames.py # ffmpeg frame extraction + auto-fps logic
│ ├── transcribe.py # VTT parsing + dedupe + Whisper orchestration
│ ├── whisper.py # Groq / OpenAI clients (pure stdlib)
│ ├── setup.py # preflight + installer [changed: local STT detection]
│ └── build-skill.sh # build dist/voxpip.skill for claude.ai upload
├── hooks/ # SessionStart status hook (Claude Code only)
├── .claude-plugin/ # plugin.json + marketplace.json [changed]
├── .codex-plugin/ # codex packaging
└── .github/workflows/ # release.yml — auto-builds voxpip.skill on tag push
Credits
Built on claude-video by @bradautomates (MIT).