OpenFilm: The Open-Source Video Agent That Turns Claude Code or Codex Into a Film Studio

OpenFilm is a free, MIT-licensed tool that turns a coding agent into a video agent: Claude Code, Codex or any agent that can run a command writes your video as a folder of web pages, and a local editor called Studio lets you change anything it made. The repo, openfilm/openfilm on GitHub, picked up 1,338 stars between October 4 and October 20, 2026, and showed about 1.3k stars and 137 forks when we read it. Its pitch is one line: "One HTML file, one function, one real editor."

If you have followed the wave of "video as code" tools on explainx.ai, this is the newest entrant, and its twist is the editor: the agent does the first draft, and you do not have to re-prompt to fix a title.

OpenFilm at a glance

How does OpenFilm work?

According to the README and the openfilm.dev site, the design is built on three ideas.

A film is a folder. The folder contains a film.html file that arranges clips on tracks. Everything stays on your disk, so you can copy, diff and version a film like source code.

A shot is a web page with a frame(t) function. Each shot exposes a function that draws any moment in time. Because it is just a page, CSS, SVG, canvas and WebGL all work, and the renderer can capture frames in parallel up to 4K. The rule that follows is determinism: given the same t, the page must draw the same picture. That is what makes rendering repeatable and lets the editor scrub to any frame.

Studio is the human half. It is a local browser editor where you move, scale and rotate page elements, edit text and styling, trim and place clips, and adjust timing and subtitles. This is the piece that distinguishes it from pure prompt-to-video flows: when the agent gets a headline wrong by two words, you click and fix it.

There is also a look command that checks rendered frames for errors, failed fonts, missing sound files and clipped text. That is a useful loop for agents: the model can run look, read the problems, and patch its own pages without a human watching.

How to try it today

The fastest path with an agent already installed:

npx skills add openfilm/openfilm

Then ask your agent for a film, for example a 30-second explainer on how a transformer attends to tokens, with a voiceover and subtitles. Without the skill, you can tell the agent to run npx openfilm, or open Studio yourself:

npx openfilm open

If you prefer a GUI first, the desktop app for macOS and Windows is available from openfilm.dev or GitHub Releases, and its chat panel can talk to Claude Code, Codex and other agents, or use its own agent with your key.

Prerequisites to check before you start: Node 22.2 or newer, ffmpeg on your path, and git if you want Studio's saved versions. Building from source needs Node 22.18, pnpm 9, ffmpeg and git. The first run downloads headless Chromium, roughly 100 MB.

What does it cost, really?

The software costs nothing and needs no account. The bill arrives only when you ask for generated assets. The README lists generated voice, music, images and video through ElevenLabs, OpenAI, Google and fal, each used with your own keys and billed by that provider. On top of that, you pay for your coding agent's tokens. A film that is mostly typography, charts and SVG animation costs little in generation fees; one built from generated clips will cost whatever those models charge. We did not run a benchmark of cost per minute, and the project does not publish one.

What can you make with it?

The site lists explainers, music videos, commercials, picture books and social ads, and points to example projects in a GitHub repository. The format suits content where structure matters: kinetic text, product walkthroughs with real UI, data-driven charts, slide-style lessons, and anything a designer would build in After Effects from vector shapes. It is weaker for photoreal footage, where it can only place clips that another tool produced. The README says so directly: output from Python, Blender or video models must come in as video or images.

How does OpenFilm compare with other agent video tools?

We have covered several approaches. They overlap, but they are not identical:

• HyperFrames from HeyGen also renders HTML to video for AI agents. The shared idea is that agents are good at writing HTML. OpenFilm's difference, per its own materials, is the bundled Studio editor and the open film.html folder standard.
• Diffusion Studio's editor-as-code makes every edit a line of code. OpenFilm keeps code as the source of truth but adds point-and-click editing on top.
• video-use and OpenMontage drive Claude Code to edit and assemble video from existing footage and tools. OpenFilm instead authors the visuals from scratch as pages.
• Our earlier look at how viral Opus 5.5 videos are made with code rather than diffusion describes the same bet from the model side: a strong coding model can draw a video if the format is code-friendly.

This table reflects what each project says about itself and our earlier coverage; we have not run a head-to-head render test.

What are the honest limitations?

• Young project. The repo had about five open issues and two open pull requests when we checked. Expect rough edges and changing formats.
• Determinism is a constraint. Anything that uses randomness or real time in a page must be seeded or driven by t, or your renders will drift. Agents will occasionally forget this; the look check helps but is not a guarantee.
• Photoreal needs outside tools. There is no built-in video model.
• Browser support. Studio works best on Chromium browsers. Safari and Firefox users should test first.
• Trademark. The code is MIT, but the OpenFilm name and logo are covered by a TRADEMARK.md. Forks should rename.
• Key hygiene. Because generation services use your own API keys, keep them out of film folders you plan to push to a public repo.
• Agent permissions. You are letting a coding agent run commands and write files on your machine. Run it in a project folder, review what it does, and consider guardrails such as AgentBeam, the agent security platform from the explainx.ai team, if you run agents with broad permissions.

A practical first-project checklist

If you want to judge OpenFilm fairly, start small and measure it yourself. Pick a 20 to 30 second piece with text, shapes and one voiceover track. Ask the agent for the film, run the look check, and open Studio to fix two things by hand, such as a title and a clip's timing. Note how many agent turns the first draft took and how long the render took on your machine. Then export an MP4 and, if you use a desktop editor, a timeline file for Final Cut Pro, Premiere Pro or DaVinci Resolve to see whether it opens cleanly. Those four observations tell you more than any star count about whether the tool fits your workflow.

What are people asking on Hacker News?

The repo was posted to Hacker News (the thread had 1 point and no comments when we checked), so there is no community testing to report yet. That is worth stating plainly: the star count shows interest, but we have not seen independent reports of render quality or reliability. Treat the claims here as the project's own until readers report back.

Why does this matter?

Two trends meet here. Coding agents have become the default way to automate creative work, and "format for agents" is turning into a design discipline: files a model can read and write reliably, with a deterministic renderer and a validation command. OpenFilm adds the third ingredient, a human editor that works on the same files, so the output is not a black box. If that pattern holds, expect more categories (slides, audio, 3D) to get the same treatment.

For learners, it is also a good teaching project: a film is a folder, a shot is a function of time, and you can read all of it. That is a smaller conceptual jump than a diffusion pipeline.

Related reading

• HyperFrames: HeyGen's HTML-to-video framework for agents
• Diffusion Studio: open-source video editor as code
• video-use: Claude Code as a video editor
• OpenMontage: agentic video production with Claude Code
• How viral Opus 5.5 videos are made with code, not diffusion
• OpenCut rewrite: plugins and headless MCP
• Sources: OpenFilm on GitHub · openfilm.dev

Details are accurate as of October 20, 2026. The project is moving quickly, so check the README for current install steps and requirements.