cinecode lets an AI coding agent build films frame by frame
A Python repo that turns Claude Code or Cursor into a film crew, writing scripts and rendering every frame as code before exporting an MP4.
GitHub's trending list on 2026-11-14 included sarpstar/cinecode, a Python repository under the MIT license whose one-line description is films made with code. It was created on 2026-11-03 and by 2026-11-14 had 2,446 stars and 420 forks, all of it gathered in 11 days. The cinecode AI film agent takes a different route from prompt-to-video tools: an agent writes the script, builds every frame as code, designs the sound and hands back an MP4.
What the repository covers
The README lists short films, YouTube videos running past ten minutes, ads, explainers, vertical social clips of 30 to 60 seconds and music videos, with finished lengths from 15 seconds to beyond ten minutes. It runs on the user's own machine. The source says it works best inside Claude Code, meaning the desktop app, the VS Code extension or the terminal, and also in Cursor, and that users have run it in other agents including Codex and Antigravity. No paid service is required. Narration can optionally use the user's own Gemini key, which the source says has a free quota, ElevenLabs, or a free Turkish voice.
How a film gets made
The agent takes a one-sentence idea, then asks a few questions about how to read that idea, what the film should be made of, duration and format, look, sound and music, with a suggested answer marked for each. It shows a plan with second-by-second beats and a director card that can be edited line by line, then renders two scenes as frames for approval before the rest is built. A draft render follows, checked for motion from the first frame, loudness, readability and repetition, with a separate viewer agent watching it cold. Notes are applied only where they land. The final comes at full size, mastered to -14 LUFS, with a credits file for the video description.
In quick mode the output is one self-contained film.html where scenes are functions of time, drawn with Canvas 2D or three.js and WebGL, with sound designed from the start and rendered offline with Web Audio in the same file. Frames come from headless Chrome, which the source says produced its own 30 second sample. Film mode renders plan by plan against eight look presets.
Two speeds, two purposes
Quick mode targets work under a minute and up to roughly three minutes, with a low resolution draft, one or two revision rounds and about three approvals. Film mode runs six questions, a roadmap, a full script, a style frame, an animatic and a render, either step by step with about six to eight approvals or in one pass with about three, and adds research plus layered audio.
The agent also asks whether the film is cinematic or an ad and product piece. Ad work gets specificity and competitor checks, on-screen name labels, interface hero shots, brand colours and typefaces, and versions in 9:16, 1:1 and 16:9 with safe areas.
What this likely means for builders
This part is inference. cinecode is not a video model. It is a workflow wrapped around code generation, so the artifact is text that can be diffed, edited and rerun, which likely makes iteration more predictable than resampling pixels. The trade-off is scope: what appears on screen is limited to what a browser can draw with Canvas, three.js and WebGL. The approval gates, the independent review pass and the notes that touch only what they mention are patterns that carry over to any long agent task.
Limits and open questions
The source gives no hardware requirements and no per-film render time beyond saying that longer films need more time and usage. It lists example films but not what they cost to produce. Editing your own footage requires Whisper for word-timed transcripts. The free Turkish voice is mentioned without saying what it is. It is also unclear how the -14 LUFS target is verified and whether the browser audio path holds up on every machine.
If you are building an app with AI
The transferable idea is the shape of the process: an agent that plans, asks for confirmation, renders, self-checks and revises only what was flagged. If you want a generated visual or a chat-driven flow inside an app of your own, Draftship's guide to adding AI to your app covers the basics. The project's own code sits on GitHub under the MIT license.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.