vidclaude

Multimodal video understanding for Claude Code. Extract frames, transcribe audio in 90+ languages, build temporal timelines — all from a single command. No API key needed.

pip install vidclaude
vidclaude video.mp4 --mode standard --verbose

What it does

Drop a video in, get structured evidence out. Claude in your conversation does the thinking.

Video File
  │
  ├─ ffmpeg ──────────► Frames (adaptive, shot-aware sampling)
  ├─ faster-whisper ──► Transcript with timestamps (large-v3, 90+ languages)
  ├─ pytesseract ─────► On-screen text / OCR (optional)
  ├─ scene detection ─► Shot boundaries
  │
  └─► Timeline ──► evidence.md ──► Claude reasons over it

Works with Hindi, English, Spanish, Japanese, Arabic — any of the 90+ languages Whisper supports. Language is auto-detected.

Prerequisites

Requirement	Install
Python 3.10+	python.org
ffmpeg	Windows: winget install ffmpeg / macOS: brew install ffmpeg / Linux: sudo apt install ffmpeg

Install

pip install vidclaude

That's it. First run downloads the Whisper model (~3GB, one time).

Usage

With Claude Code (recommended)

Set up the skill once:

vidclaude --install-skill

Then in Claude Code, just say:

"analyze the video at C:/Users/me/Videos/meeting.mp4"

"what does the speaker say about the budget in presentation.mp4?"

"when does the logo appear in intro.mov?"

Claude runs the extraction, reads the evidence report + frames, and answers your question. Your Max/Pro plan covers everything — no API key needed.

Follow-up questions about the same video are instant (cached).

From the command line

# Standard analysis — good for most videos
vidclaude video.mp4 --mode standard --verbose

# Quick — fewer frames, faster
vidclaude video.mp4 --mode quick

# Deep — dense frames, full OCR, detailed
vidclaude video.mp4 --mode deep --verbose

# Batch process a folder
vidclaude ./videos/ --verbose

# Skip audio transcription
vidclaude video.mp4 --no-audio

# Force fresh extraction (ignore cache)
vidclaude video.mp4 --no-cache

Output

Every run creates a .vidcache/ directory next to the video:

.vidcache/a3f7b2c1/
  evidence.md        ← Human-readable report (Claude reads this)
  frames/            ← Extracted JPEG frames
  transcript.json    ← Timestamped transcript
  timeline.json      ← Unified event timeline
  meta.json          ← Video metadata
  shots.json         ← Shot boundaries
  ocr.json           ← On-screen text
  summaries.json     ← Scene/chapter summaries

Modes

Mode	Frames	Whisper model	OCR	Best for
quick	~20, uniform	base	skip	Short clips, fast overview
standard	~60, shot-aware	large-v3	keyframes	General use
deep	~150, burst sampling	large-v3	all frames	Long videos, detailed analysis

Smart frame budget: If a video is too long for the frame limit, FPS is automatically reduced. Shot boundary frames are always prioritized.

CLI Reference

vidclaude [input] [options]

positional:
  input                    Video file or folder

setup:
  --install-skill          Copy SKILL.md to current dir for Claude Code

processing:
  --mode {quick,standard,deep}   Processing mode (default: standard)
  -f, --fps N              Override frames per second
  -m, --max-frames N       Override max frame count
  --no-audio               Skip audio transcription
  --no-ocr                 Skip OCR extraction
  --no-cache               Force re-extraction

output:
  --extract                Print cache path summary
  -o, --output FILE        Write output to file
  --verbose                Show detailed progress
  --batch-summary          Cross-video summary for folders

How it works

Built on a multi-layer video understanding architecture:

• Layer A — Ingestion: Validates format (MP4, MOV, MKV, WebM, AVI), extracts metadata via ffprobe
• Layer B — Segmentation: Detects shot boundaries using ffmpeg's scene change filter
• Layer C — Adaptive Sampling: Content-aware frame selection — more frames at scene transitions, smart frame budgets per mode
• Layer D — Audio: faster-whisper with large-v3 for multilingual transcription with word-level timestamps and VAD filtering
• Layer E — OCR: pytesseract text extraction from key frames (optional)
• Layer G — Timeline: Merges speech, OCR, and scene events into a single time-sorted list
• Layer I — Memory: Hierarchical summaries for longer videos (scene → chapter → global)
• Layer J — Evidence Assembly: Generates evidence.md with frame references, transcript, timeline for Claude to reason over

Intent-aware processing

The tool classifies your question and adjusts the pipeline:

Question type	Example	What happens
Description	"What happens in this video?"	Balanced extraction
Moment retrieval	"When does the person stand up?"	Prioritizes transcript + timeline
Temporal ordering	"Does X happen before Y?"	Prioritizes timeline events
Counting	"How many cars appear?"	Denser frame sampling
OCR / text	"What text is on the slide?"	Prioritizes OCR extraction
Speech	"What did they say about revenue?"	Prioritizes transcript

Optional extras

# OCR support (also needs Tesseract binary installed)
pip install pytesseract

# Standalone API mode (use outside Claude Code)
pip install anthropic
export ANTHROPIC_API_KEY=sk-...
vidclaude video.mp4 --api -q "What happens in this video?"

Examples

Analyze a meeting recording:

vidclaude meeting.mp4 --mode standard --verbose
# → 55 frames, full transcript, timeline
# → In Claude Code: "summarize the key decisions from this meeting"

Review security footage:

vidclaude ./cameras/ --mode deep --verbose
# → Batch processes all videos in the folder
# → Dense frame extraction catches fast events

Extract text from a lecture:

pip install pytesseract
vidclaude lecture.mp4 --mode standard --verbose
# → Captures slide text via OCR + speaker transcript

Quick check on a short clip:

vidclaude clip.mp4 --mode quick
# → 20 frames, fast whisper, done in seconds

Troubleshooting

Problem	Fix
ffmpeg not found	Install: winget install ffmpeg (Windows) / brew install ffmpeg (Mac)
No module named 'faster_whisper'	pip install faster-whisper
Slow first run	Normal — downloading Whisper large-v3 model (~3GB, one time)
Wrong language detected	Whisper auto-detects; usually correct. For edge cases, the transcript still captures phonetics
Large .vidcache folder	Delete it to free space: rm -rf .vidcache/
Want fresh extraction	Use --no-cache flag
OCR not working	Install pytesseract + Tesseract binary

Project structure

vidclaude/
  cli.py          Argument parsing, orchestration, caching
  models.py       Data model (VideoMeta, Frame, Shot, TranscriptChunk, etc.)
  ingest.py       Video validation + ffprobe metadata
  segment.py      Shot detection + adaptive frame sampling
  audio.py        faster-whisper transcription (large-v3)
  ocr.py          pytesseract text extraction
  intent.py       Question intent classification
  timeline.py     Temporal event merging
  memory.py       Hierarchical summaries
  reason.py       Evidence assembly + optional API mode
  util.py         ffmpeg helpers, image encoding, caching
  SKILL.md        Claude Code skill definition (bundled)

License

MIT