An open-source tool that creates short-form videos (YouTube Shorts, TikTok, Facebook Reels, 16:9 YouTube), lets you edit them by hand or with AI, and publishes them to your accounts. Everything is set up from the web UI: no .env, no Docker.

By default videos are made with Claude Motion: like Anthropic's Claude Motion, Claude designs and animates every scene as code — kinetic typography, diagrams, charts, shapes and transitions — here rendered locally with HyperFrames. A Quick engine (voice-over and captions over stock clips or animated backgrounds) is there when you need a video in a minute.

 Topic / Niche ─► Claude ─► scenes ─► edge-tts voice + stock clips ─► ffmpeg render ─► edit loop ─► YouTube / TikTok / Facebook
                 (Agent SDK)          (word timings)   (Pexels/Pixabay)  (captions burned in)  │
                                                                                         manual edits,
                                                                                         "edit with AI",
                                                                                         FrameJam frame comments

git clone https://github.com/thihodev/autoclip-studio.git && cd autoclip-studio
./start.sh          # Windows: start.cmd

That's the only command. The first run creates .venv and installs dependencies (with uv if it's installed, otherwise pip), then opens http://localhost:8000. After git pull, run ./start.sh again — it reinstalls only when requirements.txt changed. Use ./start.sh --no-browser on a server. Requires Python 3.11+.

The only required step is Claude (Connections → Claude): click Sign in with Claude, approve in the browser and paste the code back, or pick Use this computer's Claude Code login, or paste an Anthropic API key. The app never uses an account until you choose one.

Everything else is optional and can be connected at any time from the same tab:

• Voice-over: an ElevenLabs API key for natural voices (pick the voice and listen in Connections). Without it, free Microsoft voices are used.
• Claude Motion: click Install once (downloads HyperFrames, its skills and Chrome). Requires Node.js 22+ and FFmpeg.
• Stock footage: a free Pixabay key (or an existing Pexels key — Pexels has paused new keys) adds real video clips behind each scene. Without one, scenes use animated gradient backgrounds in your brand colors, and you can still upload your own clips or use brand b-roll.
• Social accounts, GitHub, FrameJam and Postiz.

Claude Motion uses noticeably more of your Claude plan per video than the Quick engine.

HyperFrames' own skill installer writes into your global agent folders (~/.claude/skills, ~/.agents/skills, …). AutoClip never lets it: it installs the skills into data/hyperframes/ with an app-owned HOME and links them into each video project, and the agent runs with a command allow/deny policy, file writes limited to the project folder, and no web access.

• Interface in English or Vietnamese, light or dark theme (switch in the sidebar; follows your browser and system by default). Translations live in static/locales/<lang>.json — add a file to contribute a language.

• Create: topic → Claude splits the voice-over into scenes and picks a footage search for each → render.

• Editor (per project):

  • edit each scene's text and footage search, reorder, add or delete scenes;
  • swap footage from search results or upload your own clip;
  • caption style (bold with a highlighted word, box, minimal, none), colors, position, voice and speed, camera drift;
  • Edit with AI: describe the change ("punchier hook, shorten scene 3, captions at the top") and Claude edits the timeline;
  • FrameJam review: send the render to FrameJam, click frames and comment, then Apply feedback;
  • every render is kept as a version (v1, v2, …) you can preview and compare.

• Projects: group videos per product or channel (/w/<id>). A project carries a website, a brand kit, instructions for Claude and context files, and every video in it uses them.

• Design from a website: when a topic or project names a domain (e.g. pickcourt.com) and no brand kit is chosen, AutoClip opens the site in headless Chrome (HyperFrames capture), lets Claude look at it and builds a brand kit — colors, fonts, logo, voice, call to action, the site's copy as knowledge, and real product images that Claude Motion features. Also available as Brand kits → Import from website.

• Live view: while a video is being made, the editor shows what Claude is doing step by step with the latest frame it looked at; elsewhere in the app a corner dock keeps every running video's progress in view.

• FrameJam on by default: AutoClip starts FrameJam if it isn't running and sends every render to its review. Turn it on or off for each video when you create it (the default is in Connections).

• Models & effort: pick the Claude model (Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1) and effort level (low → max) as defaults for scripts/edits and for Claude Motion in Connections, and override them per video or per AI edit. Each version records the model and effort that made it.

• Usage (/usage) shows the way Claude is connected now. On a Claude plan: how much of the 5-hour session and weekly limits (overall and per model) is used and when each resets — the same numbers as /usage in Claude Code, read from the Claude CLI without sending a message — plus tokens, runs and Claude time per day, model and task (no dollar amounts: a plan isn't billed per token). With an API key: what you pay, per day, model, task and run. Runs made the other way are left out; each video shows its own usage the same way.

• Context files (optional): attach articles, notes, product docs or transcripts (TXT, MD, PDF, DOCX, HTML, CSV, JSON, SRT/VTT) when creating a video, or later in the editor. Claude uses them as source material for the script and for AI edits.

• GitHub repository (optional): pick one of your repositories (connect GitHub to list private ones too) or paste any public owner/repo. Claude reads its description, README and latest release and bases the video on it; without a brand kit chosen, the repository's kit is reused or built from it first. The topic can stay empty — the repository is the topic.

• The composer remembers its last language, format, length, style, voice, engine, model, effort, FrameJam choice, publishing targets and brand kit for the next video.

• Brand kits: a design system for videos — see below.

• Ideas: Claude suggests topics for a niche, each with an opening hook.

• Autopilot: every N hours, generate ideas → render → optionally auto-publish.

• Publish to YouTube, TikTok, Facebook Pages (and Postiz channels) in one click.

A brand kit bundles everything that makes videos look and sound like yours:

Pick a kit when creating a video (Studio), in the editor (Style → Brand kit, with per-video toggles for logo, intro, outro and music) or for Autopilot.

Brand kits → Import from GitHub accepts owner/repo or a GitHub URL (optionally a branch or tag). Public repos work as they are; for private repos, connect GitHub in Connections with a fine-grained token that has read-only Contents access.

• If the repo contains autoclip.brand.json (it can also list context documents as brand knowledge) (at the root, .autoclip/brand.json or brand/autoclip.brand.json), it is applied exactly. See examples/autoclip.brand.json; asset paths are relative to the repo root.
• Otherwise Claude reads the README, package.json, Tailwind config, CSS variables and design-token files, and picks the logo, font, music and clips from the repo's files. The README is added as brand knowledge.

Sync from GitHub on a kit re-imports it from the same repo and ref. SVG logos and WOFF fonts can't be rendered by ffmpeg; the import warns you to upload a PNG or a TTF/OTF instead.

A sign-in and an API key can both be saved; switch between them in Connections → Claude without signing in again.

⚠️ Use your subscription for yourself only. Since June 15, 2026, Agent SDK usage on subscription plans draws from a monthly "Agent SDK credit", separate from your chat limits. Don't share your login with other people; for multi-user deployments, each user should connect their own account or use an API key. See the authentication docs.

Web sign-in drives claude auth login in a pseudo-terminal, so it works on macOS and Linux (on Windows, use this computer's login or an API key).

Each platform requires your own developer app. The Connections tab shows, for each platform, the exact redirect URI to register and where to paste the client ID and secret. Then click Connect and approve on the platform's own consent screen.

AI-generated content is labeled where the platform supports it (containsSyntheticMedia on YouTube, is_aigc on TikTok).

npx -y framejam start     # Node 22+, runs at http://localhost:2400

In the editor click Open in FrameJam, comment on frames in the review page, then Apply feedback. Claude maps each comment's timestamp to a scene, edits the timeline and renders a new version, which is pushed to the same FrameJam review automatically.

Everything user-facing lives in the UI and is stored in data/autoclip.db. Optional environment variables (see .env.example): HOST, PORT, DATA_DIR. Run with ./start.sh --no-browser on a server.

The app has no login of its own. Keep it on 127.0.0.1 (the default), or put it behind your own authentication before exposing it.

publish_options: youtube_visibility (public|unlisted|private), tiktok_method (DIRECT_POST|UPLOAD), tiktok_privacy, schedule_at (Postiz only).

• Background music library and ducking.
• More voices: Kokoro TTS, F5-TTS (voice cloning).
• AI-generated shots instead of stock footage: ComfyUI + Wan 2.x / LTX-Video.
• Clip long videos into shorts with faster-whisper + Claude picking the best moments.
• Fonts for non-Latin scripts (CJK, Thai, Devanagari).

MIT. The bundled Be Vietnam Pro font is under the SIL Open Font License. FrameJam and Postiz are separate projects used over their APIs and keep their own licenses.