AutoClip is an open-source app that turns a long video into short, platform-ready clips: you paste a link, pick a platform, and it renders up to 10 ranked clips with captions, a cover, a title, a description and hashtags. Cutting and rendering run on your own computer. It is trending on GitHub after gaining about 110 stars today, and it comes as a desktop app, a Docker web service, a command-line tool and an MCP server that agents can drive.
This article explains how it works, how to run it, what the README reports for time and cost, where your data goes, and what to check before you publish clips. Details come from the project README and the repository page. We did not run AutoClip ourselves, so behaviour and numbers below are the maintainers' account.
What AutoClip does
The pitch is "one link, one click." Per the README, you choose a target platform, and the app generates clips inside that platform's frame and duration rules. Each platform automatically renders up to 10 top-ranked eligible clips, with more available on demand. Each clip comes with a cover, title, description, hashtags and a ZIP publish kit.
Packaging choices are built in:
- Douyin and Xiaohongshu default to an interview layout.
- TikTok, Reels and YouTube Shorts default to a full-screen podcast layout.
- Bilibili and YouTube use landscape.
- You can switch between the two vertical layouts without changing the caption language.
- Framing follows the speaker and keeps the full frame when no person is found.
- A brand outro is on by default and can be turned off in Settings. The README says copied post text carries no AutoClip credit.
The README shows demo clips made from public interviews and podcasts, including conversations with well-known tech figures and entertainers. It adds that "source videos belong to their creators; these clips are shown for demonstration only." That caveat is important, and we return to it below.
What it costs: the README's own measurements
The maintainers published timing and cost from real runs, with a clear warning that these come from three different sources, not a controlled comparison.
| Version | Source | Output | Estimated text-model cost (CNY) |
|---|---|---|---|
| New, with captions | A 1 h 43 m English interview, targeted at Xiaohongshu | 7.5 min for 10 clips | 0.09 |
| New, no captions | A 2 h 52 m Chinese interview, targeted at Douyin | 29.5 min for 10 clips | 0.20 |
| Previous | A 2 h 06 m English podcast, targeted at TikTok | 65 min | 0.64 |
The runs were on 1 October 2026 on the same Apple Silicon Mac. The first two used the new pipeline and rendered 10 clips; the old pipeline rendered 33 clips, so inputs and outputs differ. The first run used creator captions, and the second used local Whisper base. Costs estimate usage of the qwen-plus text model at the time and exclude cloud speech recognition, AI images and publishing services. The README says your provider's bill is authoritative. Treat these as one team's measurements on one machine, not a benchmark.
Where your data goes
The README is explicit, which is useful:
- Local: cutting, framing and rendering run on your computer.
- Cloud analysis: sends relevant captions and post text to the model you choose.
- Visual understanding or reference-image generation: sends needed sample frames.
- Cloud transcription: sends audio. The README names Bailian cloud transcription (qwen-audio-3.1-asr-flash, about 0.1 yuan per hour of audio) as the fallback when local Whisper will not install or is too slow, and says the audio is uploaded.
- Publishing: finished clips upload to connected platforms only when you choose to publish.
- Telemetry: analytics and error reports can be disabled in Settings.
For private footage, choose local transcription and a local model through Ollama or LM Studio. The README lists default endpoints at http://localhost:11434/v1 for Ollama and http://localhost:1234/v1 for LM Studio, and notes that inside Docker localhost is the container, so point at a host address the container can reach.
Install: desktop, Docker or CLI
| You want to | Use | You need |
|---|---|---|
| Make clips on this computer | Desktop | macOS Apple Silicon or Windows x64 |
| Self-host or Linux | Docker | Docker and Compose v2 |
| Batch or agents | CLI and MCP | Python 3.10 or later (3.11 recommended) and FFmpeg |
Desktop. Download the installer from the repo's Releases page: a .dmg for Apple Silicon or a -setup.exe for Windows 10 or 11. Python and FFmpeg are bundled. On macOS 15 and later, double-click once; if blocked, open System Settings, Privacy and Security, and choose Open Anyway. On Windows, SmartScreen may ask you to choose More info, then Run anyway. The README treats these prompts as normal for the installer, so download only from the official Releases page and not from a mirror. There is no Intel Mac desktop build.
You then configure a model: choose a provider, enter its API key, select an analysis model, test the connection and save. The home screen has an example project that needs no API key. A first local install downloads about 220 MB, which the README estimates at about 15 minutes on a slow network.
Docker. Clone the repo, copy env.example to .env, create data, logs and uploads folders and run docker compose up -d --build. For a LAN address or custom domain, add the frontend address to AUTOCLIP_ALLOWED_ORIGINS.
CLI and MCP. Per the README, uvx autoclip-mcp and npx skills add zhouxiaoka/autoclip are the one-line installs. A command like this produces clips from a local file and a subtitle file:
autoclip produce talk.mp4 --srt talk.srt --platform douyin --portrait-style podcast --json
autoclip outputs PROJECT_ID --export-kits
Supplying --srt skips transcription. The MCP server exposes start_quick_output and get_quick_output_status, and autoclip mcp install opencode writes client configuration for OpenCode users. If you work with coding agents, this makes AutoClip a tool an agent can call, much like the skills in our posts on diagram-design and the Orkas local-first agent desktop.
Models you can use
In Settings you can pick a cloud API provider and enter a key, use an OpenAI-compatible service with a custom base URL, run Ollama, or run LM Studio. The README thanks a token platform sponsor that aggregates many hosted models, so the project works with the usual range of text and speech models. Cheaper text models are enough for ranking highlights and writing titles, which is why the per-run costs above are so low. The heavy work, video decoding and encoding, is done by FFmpeg on your machine, and local models run through Ollama or LM Studio.
Why it is trending
Long-form video to short clips is a large content-repurposing workflow, and paid tools charge by the minute. A free app with local rendering, a case library of example clips and an agent interface is an easy story to share. The repo shows a large history (1,297 commits on the page we read), an AGENTS.md and CLAUDE.md for contributors, a release checklist and a security policy, which suggests it is actively built. For how AI is changing what models can do with video itself, see our explainer on whether LLMs can watch video. If you edit video and need cleaner backgrounds before clipping, the team behind our sister site covers that in a guide to AI video background removers.
Who it suits and who it does not
AutoClip fits creators, podcast teams and course authors who already own long recordings and publish to several short-video platforms. One source can feed Douyin, Xiaohongshu, TikTok, Reels, Shorts, Bilibili and YouTube, each with its own frame and duration rules, and the publish kit saves the copy-and-paste work. It also fits developers who want a clipping step inside an agent workflow, because the CLI returns JSON and the MCP server reports status.
It suits you less if you need fine manual editing, because the README frames the editor as the place to adjust results, not the main workflow. It also suits you less on an Intel Mac, where only Docker and the CLI are options, or if your footage is sensitive and you cannot accept any cloud step. In that last case, plan for a slower local pipeline.
Risks and limits
- Copyright and platform rules. AutoClip clips whatever you feed it. The README's own demos use third-party interviews and say source videos belong to their creators. Use footage you own or have a licence to reuse, and check each platform's reupload and monetisation rules.
- Likeness and context. Short clips of public figures can strip context. Review every clip for accuracy before posting, especially anything that could be read as an endorsement.
- Cloud costs and data flow. Low text-model costs exclude transcription, images and publishing. Audio uploads when you use cloud transcription.
- Unsigned-app warnings. Expect OS prompts on first launch and install only from the official Releases page.
- Platform coverage. Desktop builds exist for Apple Silicon and Windows x64 only.
- Measured once. The cost table comes from three runs on one Mac, with different inputs. Your sources and models will vary.
- Mostly Chinese docs in places. Several guides, including CLI and MCP details, are in Chinese, with English summaries in the README.
How to try it safely
- Install the desktop app from the official Releases page and open the bundled example project, which needs no API key.
- Use your own interview or podcast, or footage you are licensed to reuse.
- For private material, choose local Whisper and a local model, and disable analytics in Settings.
- Review captions, framing and completeness on every clip before you save the publish kit.
- Run a second source to check the cost against the README's table before you automate a batch.
Follow @explainx_ai for more trending open-source tools.
Related reading
- diagram-design: an agent skill for brand-matched diagrams
- Orkas: a local-first AI team desktop for coding agents
- Can LLMs watch video? Claude and Gemini options
- Unslop: a Claude skill for fixing vibe-coded UI
- OpenSRE: an open-source agent for red CI builds
- app-store-compliance: a pre-submission guard for agent-built apps
