OpenMontage: Turn Your AI Coding Assistant Into a Full Video Production Studio

OpenMontage bills itself as the first open-source, agentic video production system, and its pitch is disarmingly simple: describe what you want in plain language, and your AI coding assistant handles the research, scripting, asset generation, editing, and final composition. It is built to be driven by an agent, not clicked through in an editor, so it works with Claude Code, Cursor, Copilot, Windsurf, and Codex, any coding assistant that can read files and run code.

Under the hood it is a serious system: more than ten production pipelines, over 100 tools, and more than 700 agent skill and production-knowledge files. That depth is what lets it go from a one-line prompt to a finished video instead of stopping at a script or a storyboard. It is a Python project, licensed under AGPL-3.0, and the repository had 76,864 stars and 10,197 forks when we checked on November 5, 2026, after gaining 3,700 stars in a single week.

The fastest way in is also the most interesting feature: paste a reference video and let the agent study it. Give it a YouTube Short, a Reel, a TikTok, or a local clip and say something like, here is a Short I love, make me something like this but about quantum computing. The agent analyzes the transcript, pacing, scenes, keyframes, and style, then comes back with two or three differentiated concepts, an honest tool path, cost estimates, and a sample before full production starts.

What you get back is not a vague prompt. You get what it keeps from the reference, the pacing, hook style, structure, and tone; what it changes, the topic, visual treatment, angle, and narration approach; what it will cost at your target duration before any asset generation begins; and what it will actually look like with the tools you currently have available. That cost-upfront answer is the part most AI video tools skip, and it is the difference between an experiment and a plan. Decide the format before you ask: a Short or Reel is vertical 9:16 and a standard YouTube video is 16:9, and the aspect ratio calculator gives the exact pixel dimensions for either if you are reframing your own clips to match.

The showreel is genuinely varied, which is the point: a cinematic sci-fi trailer produced end to end, a 60-second Pixar-style animated short about a lonely banana that cost 1.33 dollars to generate, a 54-second music-driven 3D showcase with gravity-shifting furniture and frozen car drifts, a 50-second vertical transformation film, a 60-second product film, and a 100-second documentary about salt that mixes real-world footage with narration and hand-authored motion graphics.

One distinction matters more than the rest. OpenMontage can make image-based videos, but it can also make actual motion video for free and open-source workflows: the agent builds a corpus from free stock footage and open archives, retrieves actual motion clips, edits them into a timeline, and renders a finished piece. That is not the usual trick of animating a handful of stills and calling it video. If you want that path, ask for a documentary montage, tone poem, or stock-footage collage and say explicitly to use real footage only.

Here is the part to understand before you start: the AI-generated clips use paid providers, so a video built on generated footage costs real money, while a video built on free stock footage costs nothing but time. OpenMontage is upfront about this at proposal time, showing you the cost before generation starts, and you choose the path yourself. The README's own examples put numbers on it: a 60-second animated short cost 1.33 dollars, a transformation film about 4 dollars, and a seven-world showcase about 5 dollars in generation costs.

The paid path runs through providers like fal.ai for FLUX images and Veo, Kling, and MiniMax video, plus ElevenLabs, OpenAI, Google, Runway, and HeyGen, each behind an API key you add to a .env file. The free path uses Piper TTS for narration, Archive.org, NASA, and Wikimedia Commons for footage, Pexels, Unsplash, and Pixabay for stock, and Remotion or HyperFrames for composition, all with no paid key at all. If you have a GPU, you can also unlock free local video generation with models like Wan 2.2. Every provider selection is scored across seven dimensions with an auditable decision log, so you can see why the agent picked what it picked.

Chat tells you what the agent said. Backlot shows you what the production is actually doing. It is a local board that fills itself in as the pipeline runs: stages light up, the script lands as a screenplay page, scene cards shimmer while assets generate, and every provider decision and dollar spent is on the wall. When a production starts, the agent opens it for you automatically.

The storyboard is also a real approval gate. Asset generation pauses on a scene-by-scene contact sheet, takes, prompts, per-asset cost, and quality scores, so you approve the visuals before the render, not after it is too late. Creative gates hold until you answer in chat, and when a run is done you can replay the whole production from its timestamps, scrubbable end to end. It is the closest thing to sitting in an edit suite while an agent does the work.

You need Python 3.10 or newer, FFmpeg, Node.js 18 or newer, and an AI coding assistant. Clone the repository, run make setup, open the project in your assistant, and tell it what you want in plain language, for example, make a 60-second animated explainer about how neural networks learn. If you write the narration yourself, keep in mind that 60 seconds holds only about 130 words at a typical speaking pace, and the speaking-time estimate in the Word Counter tells you whether your script fits. Or ask for the real-footage path: make a 75-second documentary montage about city life in the rain, use real footage only, no narration, elegiac tone, with music.

That is genuinely it for a first video. The agent researches your topic with live web search, generates AI images, writes and narrates the script with voice direction, finds royalty-free background music automatically, burns in word-level subtitles, and renders the final video. Before you see anything, the system runs a multi-point self-review: ffprobe validation, frame sampling, audio level analysis, delivery promise verification, and subtitle checks. Add API keys only when you want the paid providers; with zero keys you still get narration, footage, composition, subtitles, and post-production for free.

First, the license: OpenMontage is AGPL-3.0, which is stricter than MIT or Apache. If you run it as a network service or build on it, read what AGPL requires of you, because it is not a license to ignore. Second, costs are real on the paid path. The per-video numbers are small, a dollar or a few dollars, but they scale with length and the number of takes, and the approval gates exist precisely so you do not burn money on assets you have not seen.

Third, this is an agentic system, so results depend on the assistant driving it and the providers you have configured. A run with only free tools looks different from a run with Veo and Kling behind it, and the proposal step is where you find that out, before spending anything. Finally, it is a young, fast-moving project: 76,864 stars and a week of 3,700 new ones means the docs and pipelines change often, so treat the README as the current truth. What it offers today is remarkable either way: a full video production pipeline that starts from a video you love and ends with a finished piece, open source, with the cost stated upfront.