AutoClip: turning one video link into platform-ready short clips, with paid model calls behind it

One link in, a video, a cover and post copy out

AutoClip's workflow starts from a URL. You paste a link, choose the target platform, and it generates the clip itself, a cover image and the publishing caption, formatted for Douyin, Xiaohongshu, TikTok, Reels, YouTube Shorts, Bilibili or YouTube. Editing and rendering happen on your computer, and an editor is available when the automatic cut needs adjusting. The showcase shows two layouts for vertical output, an interview style and a full-screen podcast style, plus a horizontal mode that keeps the original frame for Bilibili. Sample clips run from 0:43 to 2:02 and carry subtitles in Chinese or English, cut from long interviews such as Bill Gates, Andrej Karpathy and Jensen Huang conversations. The changelog for recent releases mentions subtitle paging, framing on the speaker, and fixes for the long-video queue. The project describes itself as a tool for derivative highlight-reel edits, and each showcase clip links back to its original source.

The software is free, the model calls are billed per use

AutoClip costs nothing to download, but the clipping pipeline calls cloud language models, and those are billed per use by the provider. There is no bundled local model for this step. requirements.txt pins SDKs for OpenAI (openai==3.8.0), Google (google-genai==2.22.0) and Alibaba's DashScope, and the Docker configuration also passes through a SiliconFlow key. Defaults favor DashScope: LLM_PROVIDER falls back to dashscope and API_MODEL_NAME to qwen-plus. The consequence for a new user is that holding only an OpenAI or Gemini key is not enough; you also have to switch LLM_PROVIDER, or the first run asks a provider you have no account with. OPENAI_BASE_URL and DASHSCOPE_BASE_URL can redirect requests to a different endpoint. Budget for model spend on every long video you process, since each one goes through the provider.

Desktop, CLI and MCP ship under one version number

Releases are tagged as AutoClip Desktop and come quickly: v1.5.6 on October 9, v1.5.7 on October 24 and v1.5.8 on November 2, 2026, with the last push the same day. The desktop app is built with Tauri, which is why src-tauri/ sits in the repository. The CLI and MCP server are updated in lockstep with it, and the authoritative version comes from the CLI or from the MCP get_version call:

autoclip --version

The Python distribution is named autoclip-mcp, version 1.5.8, requiring Python 3.10 or newer. It installs two entry points, autoclip mapped to backend.cli:main and autoclip-mcp mapped to backend.mcp_server:main. Because the distribution and the server share a name, uv can fetch and start it in one step:

uvx autoclip-mcp

The wheel carries prompt templates, fonts, ONNX models, compressed lexicons, outro clips and render templates as package data.

Exact pins exist because the desktop build once shipped broken

requirements.txt pins every direct dependency with ==, and its header explains why: an unpinned yt-dlp or openai upgrade is how works-on-my-machine reports turn into broken releases. CI, the Docker image and the desktop bundle all install that same file. Transitive dependencies are left to pip so one file resolves both on Linux with Python 3.11, used by CI and Docker, and on the macOS and Windows portable runtime with Python 3.13. The pinned set is the one the Backend tests and docker-smoke jobs installed on September 6, 2026. The provider SDKs are bundled for a concrete reason: they used to come from a separate, unpinned installer, the portable desktop build shipped without them, and every provider call failed. If you install from source, run pip install -r requirements.txt on Python 3.10 or newer rather than picking packages by hand.

YouTube links depend on the yt-dlp version AutoClip pins

Video download goes through yt-dlp, pinned at yt-dlp[default]==2026.8.19. The [default] extra matters: it pulls in yt-dlp-ejs, the YouTube JavaScript challenge solver, along with brotli, mutagen and pycryptodomex. Without ejs, YouTube links fail with the error The page needs to be reloaded. A source install that drops the extra, or swaps in a plain yt-dlp, hits exactly that failure. The exact pin has a second consequence. When a video site changes in a way the pinned yt-dlp cannot handle, AutoClip cannot pick up a newer yt-dlp on its own; the fix arrives when the project bumps the pin and ships a release. Given the release pace, that wait has been short, but a pipeline that depends on a single platform should plan for occasional download failures until an update lands. AUTOCLIP_YT_SUBTITLE_LANGS controls which YouTube subtitle languages are requested.

The Docker stack publishes Redis on port 6379

docker-compose.yml runs a redis:7-alpine container and the AutoClip application, and a comment notes that four services share one image, autoclip:local, so it is built once. The app publishes port 8000 for the backend API and 3000 for the frontend, mounts ./data, ./logs and ./uploads from the host, and stores its database in SQLite at data/autoclip.db. Task queuing goes through Celery on Redis. Redis is mapped to host port 6379, and its command line in the compose file sets only --appendonly yes. On a laptop that is harmless. On a VPS or any machine reachable from a network, it exposes the queue port on the host, so remove the mapping or put a firewall in front before deploying. LLM keys reach the container through variable substitution from a .env file in the project root; the compose file notes that before this passthrough was added, the API_DASHSCOPE_API_KEY users set per DOCKER.md never reached the container.

Sentry and an analytics event catalog ship inside the package

Two packaging details are worth knowing before deploying AutoClip where data handling matters. sentry-sdk==2.69.2 is one of the pinned runtime dependencies. setup.py also overrides the build step to copy docs/analytics/event_catalog.json into every wheel as backend/core/event_catalog.json, described in the file as copying the tracking plan without keeping a second registry. Neither fact says on its own what is sent or when, so anyone with privacy requirements should read the event catalog and the backend configuration before the first run instead of assuming the tool stays silent. Upload-related settings, UPLOAD_POST_API_KEY and UPLOAD_POST_USER, are also passed into the container, and AUTOCLIP_ALLOWED_ORIGINS limits which browser origins may call the API. Leaving the origins variable empty in an exposed deployment is the kind of default to revisit.

Editorial conclusion

AutoClip suits creators and editors who repurpose long interviews or podcasts into vertical clips and are willing to pay per use for a cloud model such as DashScope, OpenAI, Gemini or SiliconFlow. It does not suit anyone who needs a fully offline pipeline or who cannot hold an API key. Before relying on it, confirm which LLM_PROVIDER your key matches, check that YouTube fetching works with the pinned yt-dlp, and close or firewall port 6379 if you run the Docker stack on a reachable machine.

Frequently asked questions

Official sources

1. License: MIT
2. Project website
3. README
4. Releases
5. zhouxiaoka/autoclip on GitHub