DEV Community

Cover image for Claude Code Can Now Direct Whole Films From Code — Meet cinecode
Shakeel Ahmed
Shakeel Ahmed

Posted on

Claude Code Can Now Direct Whole Films From Code — Meet cinecode

#ai #opensource #javascript #video

Claude Code Can Now Direct Whole Films From Code — Meet cinecode

Originally published on AI Growth Lab

cinecode turns Claude Code or Cursor into a film crew. It is a free, MIT-licensed skill kit by sarpstar on GitHub that plans a film, renders every frame in code with three.js or Canvas 2D, mixes the sound, and delivers a finished MP4. No video model, no render queue, no credits balance — just code, a browser, and ffmpeg. Created on 3 November 2026, it hit the top of GitHub trending within days, passing about 1,700 stars by 9 November and sitting at around 2.5K stars as it trends again on 15 November 2026.

That speed tells you something. Developers have been waiting for a way to make video the way they make software: deterministic, version-controlled, and repeatable. cinecode is the first open-source kit to package that idea into something you can install in minutes.

How cinecode works under the hood

cinecode is not a standalone app or a video generator. It is a skill — a folder of instructions, templates, and scripts your coding agent loads when you ask for a film. The agent becomes director, animator, and sound engineer, and the film itself is a program.

The pipeline has five stages:

  • Plan — Describe the film in one sentence. The agent asks a few questions (format, length, style, sound), then writes a second-by-second beat plan.
  • Script and storyboard — The agent drafts the script, a scene roadmap, and a style frame so the look stays consistent.
  • Frames in code — Every scene is a pure function of time. Short pieces live in one self-contained film.html using Canvas 2D; longer films use three.js/WebGL scenes with eight look presets, named camera moves, and a type layer.
  • Render — Headless Chrome captures the film frame by frame, exactly as the code describes. Nothing is hallucinated, because nothing is generated by a model.
  • Sound and mux — The soundtrack is built with Web Audio, rendered offline, mixed and mastered to -14 LUFS, then muxed with the video by ffmpeg into a finished MP4.

Before the expensive full render, cinecode shows you two scene stills and a low-res draft, with up to two revision rounds. That approval gate matters: a full film render can take hours, so you fix the look while it is still cheap.

Two modes: Quick and Film

  • Quick mode — Best for films under about three minutes: ads, explainers, and social clips. One self-contained file, a fast feedback loop, and only a few hundred MB of disk.
  • Film mode — For longer pieces, including 10-minute-plus YouTube documentaries, with the full roadmap, script, style frame, and animatic process. It needs several GB of disk per film during rendering.

A film can start in Quick mode and graduate to Film mode, with the beats becoming the roadmap. Prototype cheap, commit expensive only when the idea works.

What you can actually build

The gallery already runs from 15-second vertical ads to a five-minute documentary, and the kit targets five clear use cases:

  • Short films — Complete narrative pieces, rendered frame by frame from code.
  • YouTube videos and explainers — Documentaries and tutorials where visuals are diagrams, scenes, and motion graphics you can edit like source code.
  • Ads and product films — Repeatable, on-brand product spots you can regenerate for a new price or feature.
  • Vertical social clips — Short-form video for Reels, Shorts, and TikTok, cut to the right aspect ratio from the start.
  • Music videos — Clips cut to a track's beat grid, with visuals driven by the same timing as the audio.

The killer feature is editability. A client wants the logo bigger, the pacing faster, or the ending changed? You edit code and re-render — you do not re-roll a prompt and pray the model keeps the character's face the same. That is the difference between a demo and a production workflow.

Setup: what you need

cinecode runs on your own machine, inside an agent that can run local tools — Claude Code (desktop app, VS Code extension, or terminal) or Cursor. The regular chat at claude.ai cannot do this, because it cannot run ffmpeg or Chrome on your computer.

Requirements, straight from the project: Node.js 22+, Python 3.9+, ffmpeg, Chrome, Chromium, or Edge, and a usable GPU — without one, Chrome falls back to CPU, fine for tests but slow for full films.

Installation follows the standard skill pattern. In Claude Code, the simplest path is to tell the agent to install the skill from the repository; manually, you clone it into your skills folder — ~/.claude/skills/cinecode for Claude Code, or ~/.cursor/skills/cinecode for Cursor. Install only one copy.

The kit's guard hooks — checks that stop common mistakes as they happen — run in Claude Code only and need a recent client (2.1.295 or newer). On older versions, or in Cursor, the film still renders; the guard simply prints "GUARD OFF" and the skill's instructions and qc.py checks carry the safety net.

One honest cost note: the kit is free, but Claude Code is not. Anthropic requires a Pro, Max, Ultra, Team, Enterprise, or Console account, so "free" means no extra cost on top of the plan you already have. Long films also burn real agent usage, so budget tokens the way you would budget render time.

The honest limits

cinecode is a week old and built by one developer, so go in with your eyes open:

  • Render time is real. Film mode with many shots can take hours for a few minutes of finished video. This is a batch job, not a live preview.
  • Windows is unproven end to end. The test suite runs on Windows, but a full film had not yet been run end to end on Windows at the time of writing.
  • No video model means no photorealism. Stylised animation, motion graphics, and 3D scenes are the sweet spot; realistic human faces are not.
  • Sound is synthesised in code. There are no recorded voices or licensed music beds — the audio is designed, not sampled.
  • It is young. No tagged releases yet, and macOS coverage in the test suite is thinner than Windows and Linux.

None of that kills the idea. It just means cinecode is a tool for a specific job: programmable, repeatable video where control matters more than realism.

Why this matters for developers

The bigger story is a shift in how media gets made. When a film is code, it gets everything code gets: version control, diffs, code review, automated tests, and CI. A YouTube channel could keep its intro, lower thirds, and explainer templates in a repo and rebuild them with a commit. An agency could ship fifty localised ad variants from one script.

It is the same shift that made local AI practical: open tools, running on your own machine, with no per-minute meter running. If you have followed the local AI wave — from llama.cpp Just Hit 1.0 — Local AI Apps Finally Have a Stable Foundation to A 145B Open Model That Runs on Your Laptop: Edge0-145B Explained — cinecode is the same idea applied to video: own the pipeline, own the output.

Should you try it?

If you already use Claude Code or Cursor, try it this weekend on a 30-second clip. Run Quick mode, watch the draft, and judge the workflow before committing to a ten-minute documentary. If you do not have a paid Claude Code plan, wait: the kit costs nothing, but the agent it needs does not come free.

cinecode will not replace video models, and it will not replace a human film crew. But for developers who think in code, it opens a door that was closed until last week: your coding agent can now direct a film, and the film is a program you can read, edit, and re-run.

Sources

By Shakeel Ahmed — AI Growth Lab

Top comments (0)