What it does

html-video is a local studio and CLI that lets coding agents (Claude Code, Cursor, Aider, etc.) generate real MP4s from HTML and CSS templates. You paste a link, describe a video, or point at a repo; the agent builds a multi-frame storyboard, fills it with your content, and renders it locally via headless Chromium and ffmpeg. Apache-2.0, no per-render fees.

The interesting bit

The project treats video engines as pluggable backends behind a single render(input, ctx) adapter. Today only the Hyperframes engine (Chromium + ffmpeg) is wired up; Remotion, Motion Canvas, and Manim are on the roadmap but not yet built. The pitch is that you never learn the engine’s DSL — the agent and template system absorb that complexity. Whether that abstraction holds when those adapters actually land is the bet.

Key highlights

• 21 curated HTML/CSS templates (data viz, kinetic type, cinematic frames, etc.) with live previews in the studio gallery
• Content-graph storyboard: agent structures source material into nodes and edges, topo-sorted into frame sequences
• Source-aware generation: fetches web articles (including WeChat 公众号) and GitHub repos server-side, flattens to Markdown, builds video from actual content not canned filler
• Optional AI soundtrack via MiniMax (music + narration), mixed at export
• 13 coding agents auto-detected on PATH; switchable from studio top bar
• Single-frame fast path skips storyboard for simple template renders

Caveats

• Only the Hyperframes engine is actually implemented; the pluggable architecture is real but the other engines are “planned” or “researching”
• The AI soundtrack requires a MiniMax call; everything else runs locally but that step is external
• README notes the studio fetches sources server-side — unclear if that means a nexu-io-hosted component or local server; the WeChat example suggests some non-local infrastructure

Verdict

Worth a look if you’re already living in agent-driven workflows and want video output without subscribing to a cloud render farm. Skip it if you need production reliability today across multiple render engines — this is a promising scaffold, not a finished multi-backend system.