Contentine Create

Generate scripts, voices, avatars, and animated video — in the browser, on your own device, with no server GPU.

Quick start · What works today · Deploy · Roadmap · Contributing

---

⭐ If you'd use a tool like this, a star helps it find collaborators. Contentine Create is a working prototype built largely by one developer, now open source because it deserves a team. Several core flows are proven end-to-end (real, playable MP4s); one headline flow is gated on a licensing decision; and there's plenty of high-value work left. If that's your kind of project, see Looking for collaborators.

What it is

Contentine Create turns an idea, a script, or a URL into finished short-form content without sending your media to a server. The heavy AI runs in your browser on WebGPU (with a WASM fallback): text-to-speech, speech-to-text, small language models for scripting, talking-head avatars, and a deterministic cartoon/motion-graphics renderer. An optional PHP/MySQL backend handles sign-in, credits, and commerce — but it is never required to create.

• On-device by default. Your script, your voice recording, your render — none of it needs to leave the machine. No server GPU bill; works offline after the models are cached.
• Deterministic rendering. Video is rendered frame-accurately (seek() per frame → OffscreenCanvas → WebCodecs → Mediabunny), not screen-captured. The same project file produces the same output on any device.
• License-disciplined. Every shipped model's weights license is read and recorded in NOTICE / LICENSES.md. Anything with unmet obligations ships behind a feature flag that defaults to off.

See it in action

What works today

A flow counts as "working" only if it produces a downloaded, ffprobe-valid MP4 — never a green spinner.

Flow	Status
Avatar (talking head) → MP4	✅ Proven
Audio-to-video (upload audio → on-device transcription → captioned video) → MP4	✅ Proven
Cartoon (script → styled animated scene → render) → MP4	✅ Proven
Cartoon projects: save, reopen after a browser restart, back up/restore (.ccproj)	✅ Proven — evidence A05/A06
Cartoon scene edit by an AI assistant over MCP (inspect → patch → render → save)	✅ Proven with Claude Code on Chrome — evidence A12, bridge
Script generation (on-device SLM, with honest fallbacks)	✅ Proven
URL → extract → script → video wizard	✅ Up to the visuals step
URL → video (AI images) → MP4	⚠️ Gated on a model-license decision — see ROADMAP

Quick start

Requirement	Minimum	Notes
Node	20+	for the frontend app and the optional assistant bridge
Browser	recent Chrome / Edge	WebGPU for the fast path; WASM fallback otherwise
PHP + MySQL/MariaDB	PHP 8.0+	optional — only for sign-in, credits, and sync

git clone https://github.com/BeyondITL/contentine-create.git
cd contentine-create

# Frontend — the app itself
cd frontend && npm install && cd ..
bash scripts/setup-assets.sh      # downloads the open model weights the app uses
cd frontend && npm run dev        # → http://localhost:3000

The app runs frontend-only out of the box. To enable sign-in, credits, and sync, set up the optional backend: copy admin_config.php.template → admin_config.php (kept out of git; never commit real secrets) and run cd backend && composer install.

Project layout

frontend/      Next.js 14 static-export app — the engines, UI, and Zustand stores
backend/       Optional PHP 8 API (auth sync, credits, commerce). PSR-4, no framework
database/      SQL schema + ordered migrations
bridge/        Local MCP companion that lets an AI assistant drive a paired Cartoon Studio tab
samples/       Versioned sample projects (cartoon scene plans)
scripts/       Asset setup, model manifest/BOM checks, and verification harnesses
docs/          Engineering, UI/UX, and implementation specs (cartoon: architecture, commands, project format)
.claude/       Claude Code project config: laws, hooks, subagents, and domain skills

Architecture

• Frontend: Next.js 14 static export (output: 'export', no SSR) — every engine import is dynamic. State in Zustand stores; IndexedDB/OPFS as the primary data layer.
• Engines: Kokoro TTS (voices), Whisper/Moonshine (ASR + captions), WebLLM small language models (scripting), TalkingHead + three.js (avatars), a GSAP-timeline cartoon engine — all exporting through WebCodecs + Mediabunny in a Worker.
• Backend (optional): PHP 8 front-controller (backend/public/index.php), PSR-4 Contentine\ namespace, MySQL. Non-blocking sync/commerce only.

Full detail: ARCHITECTURE.md and docs/ENGINEERING_SPEC.md.

Deployment

It's a static frontend (frontend/out/) plus an optional PHP backend, so it deploys on almost anything — shared host, VPS, container, or a static CDN paired with any PHP host. See DEPLOYMENT.md for the full, vendor-neutral guide, with ready-to-edit configs for Apache (.htaccess), Nginx, and IIS. The one hard requirement is cross-origin isolation (COOP: same-origin + COEP: require-corp) for the in-browser engines.

Models & licenses

Contentine Create orchestrates third-party open models but does not redistribute their weights — scripts/setup-assets.sh downloads them from their official sources. Each model's license, obligations, and status are recorded in NOTICE and LICENSES.md. Notably, SD-Turbo image generation ships disabled by default pending acceptance of the Stability AI Community License (attribution + the US$1M revenue cliff). Read NOTICE before enabling any gated model.

Looking for collaborators

The honest ask: one developer can't take this to a market-grade product alone. If you care about on-device AI, WebGPU, TTS/ASR, deterministic rendering — or just want to help a real codebase grow — you're very welcome here.

Good places to start:

• Issues labeled good first issue and help wanted.
• The open items in ROADMAP.md — the URL→video visuals story, the script-parser hardening, and accessibility are all well-scoped.
• CONTRIBUTING.md — dev setup, the "no facades / evidence before done" bar, and the graphify knowledge-graph tooling this repo uses.

Contributors

Everyone who lands a change is listed on GitHub's contributor graph and credited in release notes. Want to join them? Start with CONTRIBUTING.md: comment /claim on an issue and it's yours.

License

Code is licensed under AGPL-3.0 — use, modify, and self-host freely, but if you run a modified version as a network service, you must publish your source. Third-party model weights keep their own licenses (see NOTICE). Contributions are accepted under the Contributor License Agreement.

---