io.github.huseyinstif/klaket-mcp

🎬 Klaket

Turn any video into LLM-ready data.

A klaket is a clapperboard — the tool that syncs sound and image on a film set. Klaket syncs video with LLMs.

LLMs read text. The web became readable with scrapers — but video, the largest store of human knowledge, is still locked away. Klaket unlocks it: give it a video URL or file, get back structured, timestamped, LLM-ready data.

pip install klaket
klaket ingest "https://youtube.com/watch?v=..." --wait

{
  "transcript": [
    { "start": 14.32, "end": 19.80, "speaker": "S1", "text": "So let's deploy this with docker compose..." }
  ],
  "scenes": [
    { "start": 190.0, "end": 342.5, "keyframes": ["scene_004_01.jpg"] }
  ],
  "chapters": [...],
  "summary": "..."
}

Features

• 📝 Transcript — timestamped speech-to-text in ~100 languages (auto-detected) with word-level timestamps; pick the model per job ("model": "medium")
• 🎙️ Podcasts too — pass an audio file/URL (mp3, m4a…) and Klaket skips the visual stages, deriving chapters from speech pauses
• 🗣️ Speaker diarization — who said what (S1/S2/…), local & keyless (sherpa-onnx)
• 💬 Subtitles — ready-to-use .srt / .vtt files with speaker labels
• 🎞️ Scene detection — content-aware scene boundaries + keyframes per scene
• 🔎 On-screen text (OCR) — reads slides, terminals and captions per scene, local & keyless
• 🧩 One JSON timeline — transcript, scenes, frames and on-screen text aligned on a single timeline
• 🔌 Works offline, no API key required — the core pipeline uses zero LLM calls
• 🧠 Pluggable model layer — optional scene descriptions via local VLMs (Ollama) or any OpenAI-compatible endpoint (KLAKET_VLM=off by default)
• 🤖 MCP server — let coding agents "watch" any video and find moments inside it
• 🔍 In-video search — GET /v1/jobs/{id}/search?q=… finds the exact moment
• ▶️ Playground — the dashboard plays the video with a click-to-seek, live-highlighted transcript

SDKs

# pip install klaket
from klaket import Klaket
result = Klaket().process("https://youtube.com/watch?v=...", num_speakers=2)

// npm i klaket-sdk
import { Klaket } from "klaket-sdk";
const result = await new Klaket().process("https://youtube.com/watch?v=...");

Give your agent eyes

# Claude Code
claude mcp add klaket -- npx klaket-mcp   # KLAKET_API_URL defaults to localhost:8484

Then: "Watch https://youtube.com/watch?v=… and summarize the commands the presenter runs." The agent gets klaket_ingest, klaket_job_status and klaket_get_result tools.

Quick start

git clone https://github.com/huseyinstif/klaket.git && cd klaket
docker compose up --build
# API on :8484, dashboard on :5180
curl -X POST localhost:8484/v1/ingest \
  -H "Content-Type: application/json" \
  -d '{"url": "https://youtube.com/watch?v=..."}'

That's it — no API keys, no GPUs required. make help lists developer shortcuts (make up, make test, make e2e).

Architecture

client ──► Go API ──► Redis queue ──► Python worker (ffmpeg · faster-whisper · scenedetect)
                │                          │
            dashboard ◄────────────────────┘   /data/jobs/<id>/result.json

• apps/api — Go, job orchestration
• apps/worker — Python, media pipeline
• apps/dashboard — React dashboard

Self-host vs Cloud

Klaket is open source (AGPL-3.0) and fully self-hostable. A hosted, pay-per-minute cloud API with managed GPUs is planned — join the waitlist (coming soon).

Status

🚧 v0.7 — pre-1.0, moving fast. Star the repo to follow along.

License

AGPL-3.0. SDKs and clients will be MIT.

Contact

Built by Hüseyin Tıntaş — X (@1337stif) · LinkedIn