AutoClip — guide skill for Shorts & Reels

For AI agents: find the right skill in one call — curl -s "https://varg.ai/how-to/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://varg.ai/how-to/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page. From a terminal: curl -fsSL https://varg.ai/install | sh installs the vargai CLI (https://varg.ai/how-to/cli.md): vargai search and vargai install <slug> with no key, every varg model after vargai login.

Free MIT desktop app, CLI and MCP: a link or long video in, platform-ready clips with covers and post copy out.

What it does

AutoClip by zhouxiaoka is an MIT-licensed AI clipping tool for interviews, podcasts and courses. Paste a YouTube or Bilibili link or pick a local file, choose Douyin, Xiaohongshu, TikTok, Reels, YouTube Shorts, Bilibili or YouTube, and it renders up to 10 ranked clips per platform, each with a cover, title, description, hashtags and a ZIP publish kit. Cutting, speaker framing and rendering run on your computer; captions come from the source, local Whisper or SenseVoice, or cloud speech-to-text, and the moments are picked by the model you choose, Ollama and LM Studio included. It ships as a desktop app (Apple Silicon Mac, Windows x64), a Docker web app, and a CLI with an MCP server (uvx autoclip-mcp) for agents; the main README is in Chinese, with an English one beside it.

How to use AutoClip, step by step

  1. Desktop: download the latest release, aarch64.dmg for an Apple Silicon Mac or x64-setup.exe for Windows 10/11. If macOS 15+ blocks it, allow it in System Settings → Privacy & Security → Open Anyway; at Windows SmartScreen choose More info → Run anyway.
  2. Open Settings, pick a model provider, paste its API key, test the connection and save. For a local model start Ollama (http://localhost:11434/v1) or LM Studio (http://localhost:1234/v1) and load a model first.
  3. On the empty home screen click the example project button (its label is in Chinese and means "see an example project"); it needs no key and the example ships inside the installer. Then paste your own link or file, pick the platform and a vertical layout (interview or full-screen podcast).
  4. No subtitles on the source? Local Whisper downloads about 220 MB the first time. To skip it, switch Settings → Default models → Speech to subtitles to Alibaba Bailian cloud transcription, which uploads the audio.
  5. For agents: uvx autoclip-mcp runs the MCP server. Register it in Claude Code with claude mcp add-json autoclip '{"command":"uvx","args":["autoclip-mcp"],"env":{"AUTOCLIP_API_KEY":"${AUTOCLIP_API_KEY}"}}' (the single quotes keep the key out of the config file), and add the skill with npx skills add zhouxiaoka/autoclip.
  6. Ask: "Make Shorts and TikTok clips from <YouTube link> and give me the covers and post text." The agent calls start_quick_output, polls get_quick_output_status, then asks once more with export_kits=true for the files.
  7. From a shell: autoclip produce talk.mp4 --srt talk.srt --platform douyin --portrait-style podcast --json, then autoclip outputs PROJECT_ID --export-kits.

What you get per platform

Each platform you pick gets up to 10 top-ranked clips that fit its frame and length rules, plus more on demand. Every clip comes with a cover, a title, a description, hashtags and a ZIP publish kit; copied post text carries no AutoClip credit. Douyin and Xiaohongshu default to the interview layout, TikTok, Reels and Shorts to full-screen podcast, Bilibili and YouTube stay landscape. Framing follows the speaker and keeps the full frame when nobody is found. A brand outro is on by default; switch it off in Settings.

From the CLI and MCP the default template is classic, so the same call gives the same result across upgrades; ask for --template editorial or template: "street" explicitly.

Where your data goes

Cutting and rendering never leave your computer. Cloud analysis sends captions and post text, visual analysis sends sampled frames, cloud speech-to-text sends the audio; local models and local Whisper send nothing. Clips upload only when you publish (Bilibili or Upload-Post accounts). Analytics and error reports can be turned off in Settings.

MCP in other agents

Job status always carries progress, stage, eta and poll_after_sec; cancel_job stops ffmpeg and keeps finished clips. The MCP no longer takes an api_key argument: use AUTOCLIP_API_KEY or the desktop settings.

Docker

git clone https://github.com/zhouxiaoka/autoclip.git, cd autoclip, cp env.example .env, mkdir -p data logs uploads, docker compose up -d --build, then open http://localhost:3000 (API docs on port 8000). Inside Docker localhost is the container, so point a local model at an address the container can reach.

Time and cost the README measured

On one Apple Silicon Mac in October 2026: a 1 h 43 min English talk with creator captions became 10 Xiaohongshu clips in 7.5 minutes for about CNY 0.09 of qwen-plus; a 2 h 52 min Chinese talk without captions (local Whisper base) took 29.5 minutes for about CNY 0.20. These are three different sources, not a benchmark, and exclude cloud speech-to-text and images.

Which docs are in English

README-EN, the FAQ, the install guide, the Docker guide and the OpenCode guide have English versions. The CLI/MCP guide (docs/CLI_AND_MCP.md) and the agent skill (skills/autoclip/SKILL.md) are Chinese only; this page sums them up.

Use when

Not for

Install

AutoClip: questions and answers

Is AutoClip free? Do I need an API key?

The app is free and MIT. Cloud models and cloud transcription bill you at their own rates; Ollama and LM Studio need no key but need capable hardware. Posting needs your own Bilibili or Upload-Post account.

No clips were generated.

Read the failed stage first: SUBTITLE (empty or mistimed captions, missing Whisper), ANALYZE (model connection or unparseable reply), scoring (threshold too high: lower --min-score from 0.7 to 0.5) or EXPORT (FFmpeg, disk space, permissions). autoclip doctor checks the setup. Known problems are pinned in #96.

The download stops at 5%.

Fixed in v1.5.0: real download progress, retries, fallback formats and the bundled FFmpeg path. Update and try again; a restricted video may still need its own check (#163).

Windows says "mlx-whisper only supports Apple Silicon" when installing local Whisper.

Older builds blocked the install by mistake. Update to v1.3.3 or later and install again (#141).

Early issues say it barely ran. Is it usable now?

Those complaints (#59, #26) were mostly about the Docker route, which the maintainer fixed in v1.2.1 and guarded with CI. Frequent releases followed through September and October, with Mac and Windows installers, the autoclip-mcp package on PyPI and the MCP server. AutoClip is maintained by one person, so report failures with autoclip --version, your OS, the model and redacted logs.

Is there a desktop app for an Intel Mac or Linux?

No. Desktop builds are Apple Silicon and Windows x64 only; use Docker or the CLI elsewhere.

Analysing long gameplay streams used a lot of tokens.

v1.5 cut repeated analysis calls and added a cost guard for DeepSeek's official API; the cost still depends on the source and the provider (#175). Creator captions skip transcription and are the cheapest route.

Alternatives to AutoClip

Guide written by varg from the zhouxiaoka source, checked 2026-11-13.

More Shorts & Reels skills