engine

Turn a README, a paper, a web page or a folder of clips into a finished, captioned short video, one package per platform, without leaving Claude Code.

See what it makes · Quick start · How it works · Commands · Limits

Why

Turning knowledge into short videos usually means a timeline editor, a caption tool, a thumbnail tool and a checklist for every platform's safe zones and limits. video-studio compiles instead:

• Grounded: every claim on screen cites a line in your sources. verify shows what is covered.

• Platform-ready: one dist/<platform>/ package each for TikTok, Instagram Reels, YouTube Shorts, LinkedIn and Facebook, with the video, cover, captions, post copy and a QA report. lint checks captions and text against each app's UI.

• Made to be watched: plans follow a story arc with timed "reads", each list item, step or number appears as the voice says it, and Claude reviews contact sheets of its own render before handing it over.

• Works with real footage: ingest recordings or YouTube/Vimeo/Loom links, transcribe them locally in ~99 languages, find the best short clips, keep a moving speaker in frame for vertical video, cut away to graphics while they talk, and trim pauses and filler words.

• You stay in control: Claude proposes and waits for your approval; paid voices, model downloads and screen recordings ask you through Claude Code's approval dialog and follow your policy.yaml spend limits; renders can be cancelled.

• Reproducible: video.lock pins every tool, font, renderer and asset hash. Re-renders are cached scene by scene, and diff and test catch regressions.

• Local-first: ffmpeg, system text-to-speech and local whisper.cpp. Claude writes the plan, so the plugin needs no LLM API key, and none of the features below need a paid service.

Related MCP server: MotionKit MCP Server

See what it makes

Everything below was made by the plugin itself: no hand editing.

This README as a reel. docs/media/hero.mp4 is a narrated 1080×1920 final render made from this README. It uses the animated-explainer structure, the technical style and a macOS Premium voice, and every line cites a README line. The project that makes it is examples/readme-hero:

The strips below are frames from preview renders (built-in ffmpeg renderer) made during development.

15 scene kinds. The strip shows kinetic text, a stat, a timeline, before/after, a quote, a map and a lower third, from a text-over-music reel with no voiceover:

4 style packs, on the same content. From left: the default, minimal, editorial, technical and energetic:

Languages. The Whisper paper (arXiv 2212.04356) as Hindi (top) and Japanese (bottom) versions made with localize: shaped Devanagari, Japanese line breaking and script-aware captions and covers:

Your own footage. A folder of clips cut to the beat of a bundled music bed, with text over the footage (aesthetic-broll and silent-vlog):

Quick start

Requirements: Node.js 22.13+ and a system FFmpeg with libass and libx264 (brew install ffmpeg on macOS). Nothing else to install: the engine is a single bundled file. Check your setup with /video-studio:doctor.

Inside Claude Code:

/plugin marketplace add harshil-1411/claude_plugin_video_studio
/plugin install video-studio@video-studio-marketplace

Then:

/video-studio:create README.md as a 30-second 9:16 reel for instagram and youtube-shorts

Claude reads the source, proposes a hook, a scene plan and a storyboard, and waits for your approval. It then renders a preview, looks at it, then the final, and writes the packages:

dist/
├── reel.mp4  clean-master.mp4  captions.srt  captions.vtt  cover.jpg
├── video.lock  render-manifest.json  provenance.json  video-spec.json
├── instagram/        video.mp4  cover.jpg  captions.*  post.json  qa.json
└── youtube-shorts/   …

• Voice: macOS picks your best installed voice (add a Premium voice in System Settings → Accessibility → Spoken Content). An ElevenLabs key (/plugin → video-studio → Configure, stored in the OS credential store) is used only when your policy.yaml allows it or you ask for it.

• HyperFrames renderer: richer motion graphics; needs Google Chrome and a one-time install (the render skill gives the command).

• Video URLs: brew install yt-dlp to ingest YouTube, Vimeo or Loom videos (subtitles become the transcript).

• Transcription: whisper.cpp (brew install whisper-cpp); models are downloaded only after you approve.

git clone https://github.com/harshil-1411/claude_plugin_video_studio.git video-studio
cd video-studio && pnpm install
claude --plugin-dir .

How it works

flowchart LR
  A["Sources<br/>md · pdf · docx · pptx · url · repo<br/>video · audio · clip folders"] --> B["ingest<br/>ContentIR + evidence refs"]
  B --> C["plan<br/>brief · grounded VideoSpec · storyboard"]
  C -->|your approval| D["render<br/>voice · scenes · captions · music"]
  D --> E["lint + QA<br/>platform contracts · WCAG · loudness"]
  E --> F["dist/&lt;platform&gt;/<br/>video · cover · captions · post · qa"]

• Claude is the creative engine. Skills guide Claude to write the brief and the scene spec.

• The engine does the rest. A bundled MCP server validates, renders, runs QA and packages. It is deterministic, cached and has no LLM calls.

• Three contracts connect the stages:

  • ContentIR: the sources, with provenance.

  • VideoSpec: a provider-neutral scene graph.

  • RenderManifest: exactly what happened.

  Their JSON Schemas are in schemas/.

• Platform facts are data, not code: platform-specs/*.yaml record each app's limits and UI masks, with a source URL and the date they were verified.

What you can make

Examples

Commands

What it does not do (yet)

• No generative video or avatars yet. Runway, HeyGen and fal.ai adapters are planned (Phase 7, needs keys). Until then those scenes render as titled placeholder cards. Sora is intentionally not supported.

• No posting or analytics. It produces packages and post copy; you upload them. Platform "trending sounds" are added in each app, and post.json reminds you of that.

• HyperFrames is optional. The built-in ffmpeg renderer covers every scene kind. The richer HyperFrames renderer needs its own install and Google Chrome.

• Whisper models are downloaded only with your consent (about 148 MB; 488 MB for the speaker-turn model). You can supply SRT/VTT captions instead.

• Speaker turns are English-only and label two alternating speakers (S1/S2); rename them if there are more.

• Subject tracking is automatic on macOS only. Elsewhere Claude marks the subject from shot sheets.

• No colour grading yet. Dark or bright footage gets a warning, not a fix.

• Demo capture never starts your app. You start it and give the URL, and every step is approved first.

The roadmap is in docs/PLAN.md and the current state in docs/HANDOFF.md.

pnpm install
pnpm typecheck        # tsc -b
pnpm test             # vitest
pnpm schemas          # regenerate schemas/*.schema.json
pnpm bundle           # build dist/mcp.mjs (single-file ESM, committed)
pnpm smoke            # start dist/mcp.mjs over stdio and check its tools
claude plugin validate --strict .claude-plugin/plugin.json   # plugin + skills
claude plugin validate --strict .                            # marketplace
pnpm check            # all of the above plus golden frames, stopping at the first failure
pnpm hooks            # once: run pnpm check before every git push

After changing anything under packages/, rerun pnpm bundle and commit dist/mcp.mjs (pnpm check fails if the committed bundle is stale).

Data lives next to the code: skills/, templates/, styles/, music/, fonts/ and platform-specs/.

Contributing

Guides for adding archetypes, style packs, platform packs and providers are in docs/contributing/. Ingested content is always treated as untrusted data: the plugin never executes code from sources.

License

Apache-2.0. See LICENSE. The bundled fonts are OFL-1.1 (see fonts/README.md), and the bundled music beds are CC0 (see music/README.md).