OpenMontage at 65K Stars: One Sentence to a Video, but Revisions Are the Real Barrier

An open-source project called OpenMontage once appeared at the top of GitHub Trending, and it has already reached 65K+ stars. What it does is very specific: turn AI coding assistants like Claude Code and Cursor into a video production pipeline. You describe your needs in natural language, and it researches, writes a script, finds assets, does voiceover, generates subtitles, and renders the final output.

Most AI video tools stop at giving you one prompt and one shot. OpenMontage wants to cover the entire production workflow: research, proposal, script, scene plan, assets, editing, and compositing. The repository lists more than 10 pipelines, 100+ tools, 700+ agent skill files, and support for 60+ providers. It can also cut real videos from free or open assets, not just animate still images.

The original tweet gave several cost examples: a 60-second Pixar-style animated short, THE LAST BANANA, with plot, camera movement, piano BGM, and dynamic subtitles, cost $1.33 in total; a brain-computer interface product promo cost less than $1. More samples are in the repo and on the YouTube channel, including a sci-fi trailer, a 3D object showcase, and a documentary about the history of salt. Each video comes with the full prompt, pipeline, and cost for easy reproduction.

The core pain point it addresses is straightforward: ordinary people do not lack a particular generation model; they lack a workflow that connects research, script, assets, voiceover, subtitles, and editing. OpenMontage breaks this workflow into pipelines, with a director skill for each stage that teaches the agent how to execute, how to self-check, and when to stop and wait for human confirmation.

The Backlot dashboard lays out the script, scene cards, asset generation, and costs on a local page. Storyboards can also serve as approval gates: before asset generation, you review a contact sheet containing takes, prompts, per-asset cost, and quality scores. The project also includes budget control; the default total budget cap is $10, and any single operation over $0.50 will pause for confirmation.

Using it is not complicated. Prepare Python 3.10+, FFmpeg, and Node.js 18+, then install one of Claude Code, Cursor, Copilot, Windsurf, or Codex. After cloning the repo, run make setup and describe your needs in the AI coding assistant. It can run with zero API keys: Piper handles offline voiceover, Archive.org, NASA, and Wikimedia Commons provide open assets, Remotion and HyperFrames handle compositing, and FFmpeg finishes up. To improve visual quality, connect paid providers such as Kling, Veo, Runway, Suno, and ElevenLabs.

Reference video breakdown is another practical feature. Drop in a YouTube video, Short, Reel, TikTok, or local clip, and it analyzes the transcript, pacing, scenes, keyframes, and style, then gives 2 to 3 differentiated concepts, tool paths, cost estimates, and sample clips. One user who got it running said the chain works end to end, and breaking down the script pacing of viral videos is the most useful part. But they also warned that some intermediate steps cannot be fully automated and require human help, so do not expect to let go completely.

Another user's concern is more specific: when revising, can you change just one part, such as changing only the camera movement without rerunning the whole video? If every revision reruns the entire piece, $1.33 looks cheap, but three revisions are not cheap. The materials do not directly answer the granularity of revisions. The project has checkpoints, human approvals, and decision logs, and the workflow is resumable; but whether changing only camera movement requires rerunning the whole video depends on the specific pipeline and toolchain, so you have to actually run it once.

Some people also complain that there are too many features and they do not know where to start. That feedback is not surprising. With more than 10 pipelines and 100+ tools laid out together, it is indeed easy to be confused at first glance. A safer approach is to first pick a 30-second short, get the zero-API-key path working, and then decide whether to connect paid models.

OpenMontage is worth trying, but it moves the editing barrier to review and revision. One sentence to a video is the entry point; revising it until it is publishable is the real workload. The open-source license is AGPLv3, so read the terms carefully before commercial use. The repo is here: OpenMontage, and samples are on YouTube.