Skip to content

← The Lab

AutoClip: an open-source clipper that turns long interviews and podcasts into Shorts with captions, covers and post copy, with a CLI, MCP and agent skill

AutoClip is an MIT desktop app (with Docker, CLI and MCP versions) that takes one long video and returns up to 10 platform-ready clips, each with a cover, title, description, hashtags and a publish kit. Cutting, speaker framing and rendering run on your computer, and you choose which model does the analysis, including local ones. It showed up on GitHub's daily trending list on Nov 19.

What it does. Paste a link or drop in a file, pick a platform (Douyin, Xiaohongshu, TikTok, Reels, YouTube Shorts, Bilibili or YouTube), and AutoClip finds the strongest moments within that platform's length and frame rules. It renders the top-ranked clips with paged captions, and you can generate more on demand. Vertical clips come in an interview layout or a full-screen podcast layout, framing follows whoever is speaking, and each clip ships with a cover and post copy in a ZIP. A brand outro is on by default; you can turn it off in Settings.

Where the AI runs. The analysis model is your choice: a cloud API with your key, any OpenAI-compatible service, or local Ollama / LM Studio. For transcription it uses the creator's own captions if they exist, otherwise local Whisper or SenseVoice, or a cloud speech-to-text service. Per the README, cloud analysis sends captions and text, visual checks send sample frames, and cloud transcription sends audio. Cutting and rendering never leave your machine.

Agent hooks. A CLI (autoclip produce talk.mp4 --platform ... --json) and an MCP server (uvx autoclip-mcp, with start_quick_output / get_quick_output_status tools) share the desktop pipeline, and npx skills add zhouxiaoka/autoclip installs the agent skill. That makes it easy to put in a batch job or an agent workflow that also writes the posts.

The creator's numbers, with their caveats. The README lists runs from development on one Apple Silicon Mac, not a controlled comparison: a 1h43m English interview with existing captions became 10 clips in 7.5 minutes for about ¥0.09 of text-model usage, and a 2h52m Chinese talk without captions took 29.5 minutes (local Whisper base) for about ¥0.20. Cloud transcription, AI images and publishing aren't included.

Before you use it. Desktop builds cover Apple Silicon macOS and Windows x64. Intel Macs and Linux use Docker or the CLI (Python 3.10+, FFmpeg). The builds are unsigned or not notarized, so expect OS warnings, and parts of the docs are Chinese-only. The demo reel clips well-known interviews; only clip footage you own or have the rights to repurpose. The repo has existed since mid-2025 and had about 11,000 stars, with v1.5.8 out on Nov 2.

Signals: GitHub Trending on Nov 19 (#16 across all languages and #7 in Python, about 110 stars that day). No tracked-curator or official-account pickup, so not HOT.

Related tools on ReelStack

Skills & repos in our directory

More from the Lab

Topics: Video MCP servers