kavenio-youtube-transcribe 0.1.0

Turn YouTube videos into local, agent-ready transcripts.

YouTube Transcribe

A local-first CLI that turns a YouTube video into a clean, agent-ready transcript.

YouTube Transcribe downloads the video's native audio with yt-dlp, transcribes it locally with faster-whisper, and writes a portable Markdown file. It can also preserve word-level timestamps, WebVTT subtitles, and the original audio.

• No transcription API key or per-minute fee
• Local speech-to-text on macOS, Linux, and Windows
• Stable Markdown and versioned JSON outputs
• Optional transcript cleanup with an existing Codex CLI or Claude Code login
• Agent Skill included for Codex, Claude Code, and compatible tools

Install

Requirements:

• Python 3.10 or newer
• Enough disk space and memory for the selected Whisper model

Install the latest stable release, its dependencies, and the Agent Skill:

curl -fsSL https://raw.githubusercontent.com/kaveniohq/youtube-transcribe/main/install.sh | bash

The installer creates an isolated environment, exposes the command under ~/.local/bin, installs the skill for Codex and Claude Code, and runs youtube-transcribe doctor. If ~/.local/bin is not on your PATH, it prints the exact command to add it.

Prefer installing manually? Use pipx:

pipx install kavenio-youtube-transcribe
youtube-transcribe install-skill
youtube-transcribe doctor

Quick start

Transcribe a video:

youtube-transcribe "https://www.youtube.com/watch?v=VIDEO_ID"

The first run downloads the selected Whisper model. The default base model is a practical speed/quality compromise.

Usage

What you want	Command	Result
Create a transcript	youtube-transcribe "URL"	transcript.md
Choose an output location	youtube-transcribe "URL" --output-dir ./transcripts	One video folder inside ./transcripts
Add word and segment timestamps	youtube-transcribe "URL" --timestamps	Markdown, versioned JSON, and WebVTT
Keep the downloaded audio	youtube-transcribe "URL" --keep-audio	Transcript plus the native WebM/M4A audio
Refine with Codex CLI	youtube-transcribe "URL" --codex	Locally transcribed, then text-edited with Codex
Refine with Claude Code	youtube-transcribe "URL" --claude	Locally transcribed, then text-edited with Claude
Use a faster model	youtube-transcribe "URL" --model tiny	Faster draft with lower accuracy
Use a more accurate model	youtube-transcribe "URL" --model small	Slower transcript with higher accuracy
Fix the spoken language	youtube-transcribe "URL" --language en	Skips language auto-detection
Use NVIDIA CUDA	youtube-transcribe "URL" --device cuda	GPU inference with supported CUDA setup
Use browser authentication	youtube-transcribe "URL" --cookies-from-browser chrome	Lets yt-dlp read the selected local browser session
Protect existing artifacts	youtube-transcribe "URL" --no-overwrite	Fails instead of replacing output files
Install the Agent Skill	youtube-transcribe install-skill	Adds the skill to Codex and Claude discovery locations
Check the installation	youtube-transcribe doctor	Dependency and update status

Options can be combined:

youtube-transcribe "URL" \
  --output-dir ./transcripts \
  --model base \
  --timestamps \
  --codex

Run youtube-transcribe --help for the complete command reference.

Output

Each run creates a deterministic folder named from the video title and ID:

steam-machine-8cDatzBIufQ/
└── transcript.md

With --timestamps:

steam-machine-8cDatzBIufQ/
├── transcript.md
├── transcript.json
└── timestamps.vtt

Add --keep-audio to retain the native audio file in the same folder.

Artifact	Created when	Purpose
transcript.md	Always	Readable transcript with source provenance
transcript.json	--timestamps	Versioned machine contract with raw segment and word timing
timestamps.vtt	--timestamps	Subtitles for players and tools that support WebVTT
audio.webm or audio.m4a	--keep-audio	Original native audio selected by yt-dlp

The CLI writes progress to stderr and final artifact paths to stdout, which makes it safe to use in scripts and agent workflows:

[download]    YouTube audio           [████████████████████] 100%
[model]       Download Whisper base   [████████░░░░░░░░░░░░]  40%
[transcribe]  Speech to text          [██████████░░░░░░░░░░]  50%
[write]       Agent-ready artifacts

Transcript format

Markdown begins with YAML-compatible provenance followed by the complete transcript:

---
transcript_schema: 1
title: "Example video"
source: "https://www.youtube.com/watch?v=..."
video_id: "..."
language: "en"
transcription_model: "base"
edited_by: "none"
transcript_data: null
timestamps: null
---

# Example video

Complete transcript...

With --timestamps, transcript.json is the canonical machine-readable artifact. It keeps immutable raw Whisper timing separate from optionally edited prose. The normative schema is transcript-v1.schema.json.

Optional AI editing

No editor is selected by default. --codex and --claude deliberately start a separate command-line editor process after local transcription; they may use the allowance associated with your existing login.

The editor receives transcript text through standard input and cannot change the raw timestamp data:

youtube-transcribe "URL" --codex --timestamps
youtube-transcribe "URL" --claude --timestamps

Codex runs ephemerally in a read-only sandbox without user execution rules. Claude runs without built-in tools, MCP servers, slash commands, Chrome integration, or session persistence. Both run from an empty temporary directory.

Use with AI agents

The repository includes a portable Agent Skill that teaches compatible agents when to run the CLI, which options to select, and which generated file to read.

The CLI package bundles this skill. Install it with:

youtube-transcribe install-skill

This works on Linux, macOS, and Windows and writes:

~/.agents/skills/youtube-transcribe  # Codex and compatible agents
~/.claude/skills/youtube-transcribe  # Claude Code

Use --target codex or --target claude to install only one copy. Rerunning the command safely updates an unchanged managed skill. Locally modified files are preserved unless you explicitly pass --force. Run it again after upgrading the CLI to install the bundled skill version.

Restart an already-running agent, then ask naturally:

Transcribe https://www.youtube.com/watch?v=VIDEO_ID
Transcribe this video with timestamps: https://youtu.be/VIDEO_ID
Use $youtube-transcribe to turn this video into notes: https://youtu.be/VIDEO_ID
Use /youtube-transcribe to create subtitles for https://youtu.be/VIDEO_ID

The active agent does not need --codex or --claude; those options start a second editor process.

Local development

git clone https://github.com/kaveniohq/youtube-transcribe.git
cd youtube-transcribe
python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
pytest

On Windows PowerShell, activate the environment with:

.venv\Scripts\Activate.ps1

Before submitting a change, run:

ruff check src tests
ruff format --check src tests
pytest --cov=youtube_transcript --cov-fail-under=80
python -m build
python -m twine check dist/*

Development references:

• AGENT.md — setup, architecture, invariants, and agent-facing implementation guidance
• CONTRIBUTING.md — contribution and verification rules
• ARCHITECTURE.md — detailed contracts, design decisions, and implementation evidence
• RELEASING.md — versioning and release procedure

Default tests are offline: they do not download videos or models and do not invoke Codex or Claude.

How it works

YouTube URL
  -> yt-dlp native audio download
  -> local faster-whisper transcription
  -> optional Codex or Claude text editing
  -> atomic Markdown / JSON / WebVTT output

Provider-specific code stays behind small adapters. Temporary downloads are cleaned automatically, and final artifacts are written only after the requested pipeline succeeds.

Current limitations

• One video per invocation; playlists and channels are not supported.
• Transcripts over 600,000 characters are not sent to an optional editor.
• Live streams, members-only videos, regional restrictions, and YouTube bot challenges depend on yt-dlp and may require browser cookies.
• The first transcription for a model requires a model download.

The package supplies yt-dlp's supported Deno runtime and uses faster-whisper's PyAV decoder. A global JavaScript runtime, FFmpeg, and FFprobe are not required.

Updates

After a successful interactive run, the CLI checks PyPI at most once every seven days and prints a notice when a newer version is available. It never installs updates, changes stdout, or sends transcript/source metadata.

Set YOUTUBE_TRANSCRIBE_NO_UPDATE_CHECK=1 to disable update checks.

Contributing

Issues and focused pull requests are welcome. Please read CONTRIBUTING.md before making a change and include tests for behavior changes.

License

Licensed under the MIT License.