OpenMontage Video Studio

TL;DR OpenMontage is a free, open source project on GitHub that turns Claude Code, Cursor, or any AI coding assistant into a full video production system. You describe a video in plain English and the agent handles research, script, visuals, voice, music, subtitles, and the final render on its own. There is a completely free path using open archives and free narration, and a paid path using premium image and video generators where most videos cost somewhere between two cents and a few dollars. This guide covers the setup, the exact prompts to run, how the system works, and how to pull the best possible output from it.

What this actually is

Most AI video tools give you one clip from one prompt. You still have to write the script yourself, find music, add captions, and stitch everything together by hand.

OpenMontage skips that. It is a project you install once, and from then on your AI coding assistant, Claude Code, Cursor, Copilot, Windsurf, or Codex, becomes the whole production team. You give it a one line brief. It researches the topic with live web search, writes the script, generates or sources the visuals, records the narration, finds royalty free music, burns in the subtitles, and renders the finished file.

It went to number one on GitHub Trending within days of release and has passed 34,000 stars.

This is open source software you run yourself, not a hosted app. You will need to be comfortable opening a terminal. If that is not you, skip to the honest limits section before you start.

What you will need

Python 3.10 or newer. Get it at python.org if you do not already have it.

FFmpeg. Install with brew install ffmpeg on Mac, sudo apt install ffmpeg on Linux, or download it from ffmpeg.org on Windows.

Node.js 18 or newer. Get it at nodejs.org.

An AI coding assistant. Claude Code, Cursor, GitHub Copilot, Windsurf, or Codex. The project ships instruction files for all five.

That is the full requirement list. No paid subscription is required to get started.

The setup

Step 1. Clone the project and run setup

git clone https://github.com/calesthio/OpenMontage.git
cd OpenMontage
make setup

If your machine does not have make, run this instead on Mac or Linux:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
cd remotion-composer && npm install && cd ..
python -m pip install piper-tts
cp .env.example .env

Step 2. Open the folder in your AI coding assistant Open the OpenMontage folder in Claude Code, Cursor, or whichever assistant you use. The project already includes a guide file written specifically for that assistant, so it knows how to operate the system without any extra instructions from you.

Step 3. Describe the video you want Type your request directly into the assistant, in plain English. No special syntax, no menus.

Example, zero setup cost

Make a 60-second animated explainer about how neural networks learn

Example, real footage path, also free

Make a 75-second documentary montage about city life in the rain.
Use real footage only, no narration, elegiac tone, with music.

Step 4. Add provider keys, only if you want premium generation This step is optional. If you want AI generated video clips or premium voices instead of the free path, open the .env file and add whichever keys you have. Every key is optional. Add none and the free path still works fully.

FAL_KEY=your-key            # FLUX images plus Veo, Kling, MiniMax video
PEXELS_API_KEY=your-key     # free stock footage, free to get
PIXABAY_API_KEY=your-key    # free stock footage, free to get
ELEVENLABS_API_KEY=your-key # premium voice narration
OPENAI_API_KEY=your-key     # OpenAI narration and images
GOOGLE_API_KEY=your-key     # Google narration, 700 plus voices

The two paths

Real examples from the project's own test videos: a documentary style piece cost 2 cents. A product ad using a single provider cost 69 cents. A Pixar style animated short with six generated video clips, narration, music, and captions cost 1 dollar and 33 cents.

The production types it can run

Starting from a video you already like

Instead of writing a prompt from nothing, you can hand the assistant a video you already like, a YouTube short, a Reel, a TikTok, or a file on your own computer, and ask for something similar on a different topic.

Here's a YouTube short I love. Make me something like this, but about
CRISPR for high school students.

The assistant reads the reference video's pacing, hook, and structure, then tells you what it will keep, what it will change, what it will cost, and gives you a sample before running the full production.

More prompts to try

Free real footage path

Create a 60-second Adam Curtis style archival collage about 1950s consumer
optimism. Prefer Archive.org and Wikimedia footage.

With one image provider configured, roughly 15 cents to 1.50 dollars

Create a 30-second Ghibli style animated video of a floating library in the
clouds at golden hour

Full setup with video generation, roughly 1 to 3 dollars

Create a cinematic 30-second trailer for a sci-fi concept: humanity receives
a warning from 1000 years in the future

How the system works, so you can steer it

The reason output quality is in your hands is that this is not one black box. Every video runs through the same stages a real production team would follow, in this order: research, script, scene plan, assets, edit, then final render. You can see each stage happen.

The important part for you is where it pauses. The system stops and asks for your approval on the creative stages, the idea, the script, and the scene plan, and then runs the technical stages, the asset generation, edit, and render, on its own. Those pause points are your steering wheel, and they come before anything expensive has been generated.

Under the hood it reads three kinds of plain files you can open and change yourself: the tool list of what it can do, the pipeline playbook that defines the stages and quality checks for each video type, and the skill files that hold the know how for using each tool well. Every choice it makes, which provider, which voice, which style, gets written to a decision log you can read after the fact.

Bottom line: you do not need to touch code to control the output. You control it by giving a sharp brief, approving or rewriting at the pause points, and setting a style once. The rest of this guide is how to do exactly that.

Getting the best output

1. Write the brief like a creative director, not a search query This is the single biggest lever. A one line brief gets you a one line result, which is to say generic. Name six things every time: the format and length, who it is for, the tone, the pacing, where the footage should come from, and the platform you are posting to. The more specific you are, the closer the first render lands to what you actually pictured.

Weak brief

`` Make a video about AI agents. ``

Strong brief

`` Make a 45 second explainer for founders on why AI agents fail in production. Confident, plain spoken tone. Fast cuts. Real stock footage, no cartoon look. Word by word captions, vertical for Reels. Match the pacing of this video: [link]. ``

1. Start from a reference whenever you can The reference workflow is the fastest route to a format that already works. Paste a short, Reel, or TikTok you like, or a file of your own, and ask for the same treatment on your topic. The agent studies the pacing, the hook, and the structure, then hands you two or three directions with cost estimates before it spends anything. The strongest move is to feed it your own best performing video, so the output inherits a format your audience has already told you it responds to.
2. Steer at the checkpoints. Do not rubber stamp the script It stops for your approval on the idea, the script, and the scene plan. This is where you rewrite a flat hook, tighten the pacing, or cut a weak line. Fixing the script here costs nothing, because no footage or voice has been generated yet. Fixing it after the render means paying to produce everything twice. Treat the script approval as the most valuable thirty seconds of the whole job.
3. Encode your taste in a style playbook This is what separates your videos from the wave of identical output that everyone cloning this repo is about to publish. A style playbook is a short file that locks your typography, your colors, your motion feel, and your audio profile, so every video comes out looking like yours instead of the default. It ships with a few ready made looks, such as Clean Professional, Flat Motion Graphics, and Minimalist Diagram, and you can write your own once and reuse it forever. Set this up a single time with your brand and the generic look problem goes away for good.
4. Spend where it moves the needle, nowhere else The free path is genuinely strong for documentary and explainer formats. If you add one paid thing, add a single image key, FLUX through fal.ai, for a clear jump in visual quality at close to no cost. Save premium video generation, Veo and Kling, for cinematic pieces and ads where real motion matters. The system scores every provider for each scene and logs why it chose one, so you can always see where your money went.
5. Iterate on the weak scene, not the whole video If one shot is off or the narration is flat, regenerate that one piece instead of re running the entire production. The run is saved and resumable, so you fix the ten seconds that are wrong and keep the ninety that already work. This is where the real time and cost savings live once you are producing regularly.
6. Preview before you commit the final render The system opens a live preview you can scrub in your browser before the final export runs. Catch pacing and caption timing there, where changes are free. You can also share that preview link so a partner or an editor like Renzo can review and leave notes before you spend anything on the final file.
7. Read the decision log when something looks off Every provider, style, voice, and music choice is written down with the reasoning behind it. When a video comes out wrong, the log tells you which provider or which style caused it, so your next run is a targeted fix rather than a guess. Over a few videos this is how you learn which settings consistently give you your look.

Provider cheat sheet

One key shortcut: a single fal.ai key gives you FLUX images plus Veo, Kling, and MiniMax video, so you can cover most premium generation with one signup instead of five.

Pick the right pipeline for the job

Forcing the wrong pipeline is the most common reason output disappoints. Match your goal to the right one, or just tell the agent your goal in plain words and let it recommend.

Name your target platform in the brief, YouTube, Shorts, Reels, TikTok, LinkedIn, or 21:9 cinematic, and it exports at the right size the first time instead of making you reformat afterward.

Honest limits

This is a developer tool, not a hosted app. You are running it on your own machine from a terminal. If you have never used one, expect a learning curve on day one.

The free path uses offline narration. It sounds fine but not as polished as premium voice providers. If voice quality matters most to you, budget for a narration key.

Quality still depends on your brief. A vague prompt produces a vague video. The more specific you are about tone, pacing, and what footage to use, the better the output matches what you actually wanted.

It is released under the AGPL 3.0 license. Free to use and modify, but if you distribute a modified version you are required to release your changes under the same license. Worth knowing if you plan to build a product on top of it.

Local LLM support is not available yet. You currently need a cloud based AI coding assistant to drive it. Support for fully local models is listed as planned but not shipped.

If something goes wrong

npm install fails on Windows with an argument type error. Run npx --yes npm install instead.

The assistant seems unsure how to use the project. Point it directly at AGENT_GUIDE.md in the project folder and ask it to read that file before continuing. This file is the instruction set the whole system is built around.

You added API keys but the assistant is still using the free path. Check that the keys are saved inside the .env file, not just set in your terminal session, and confirm there are no extra spaces around the equals sign.

A render looks like a slideshow of still images instead of real motion. The system has a built in check for exactly this and is meant to block it before you see the final file. If it still happens, ask the assistant to confirm which render engine was used and whether a video generation provider was actually available for that request.

The output looks generic or off brand. Set up a style playbook, or point the agent at one of the built in looks, before you generate. A good playbook is the difference between your look and the default look, and it is the fix for almost every complaint about output feeling flat.

What this changes

Making a video used to mean a script, a separate tool for voice, a separate tool for footage, and a separate tool for editing, with your own time spent gluing all of it together in between.

This collapses that into one conversation. You describe what you want once, and the same assistant you already use for other work does the research, the writing, the sourcing, and the final edit. The free path costs nothing and the paid path costs cents to a few dollars, not the price of hiring anyone.

The people who figure this out first are going to publish more, test more ideas, and spend far less time per video than the people still doing it the old way.