Run the open-source AI video studio, set up by Claude

The repo: github.com/Anil-matcha/Open-Generative-AI. That exact one, built by Anil Matcha, MIT licensed, sitting at over 26,000 stars. There are copycat forks with near-identical names, so check the username before you download anything.

What you are installing

It is a desktop studio that puts hundreds of image, video, lip sync and audio models behind one interface, instead of you juggling six different sites. The README's own model table counts 70+ text-to-image, 70+ image-to-image, 85+ text-to-video, 120+ image-to-video, 15 lip sync and 15+ audio models, and names the ones you have heard of: Flux, Kling, Veo, Sora, Seedream, Seedance, Wan and Nano Banana 2.

Route 1: no terminal, nothing to build

Start here if you have never touched a repo. The maintainers publish prebuilt installers, so there is no Node, no cloning, no commands.

1. ⬇️ Download the installer. Go to the releases page and take the file for your machine: the arm64 DMG for Apple Silicon Macs, the plain DMG for Intel Macs, the .exe for Windows, or the .AppImage or .deb for Linux.
2. 🍎 On Mac, get past Gatekeeper once. The build is not notarised by Apple, so macOS blocks it on first launch. Drag the app to Applications, then either run xattr -cr "/Applications/Open Generative AI.app" in Terminal and right click the app and choose Open, or go to System Settings, Privacy and Security, find "Open Generative AI was blocked" and click Open Anyway. One time only.
3. 🪟 On Windows, click through SmartScreen once. The installer is not code signed, so SmartScreen warns you. Click More info, then Run anyway. It installs to your local app data folder with a Start Menu shortcut.
4. 🔑 Create a MuAPI access key. The cloud models run through MuAPI. Make an account, generate a key at muapi.ai/access-keys, and copy the key value, not the key name or label (this trips up a lot of people). Paste it when the app asks. The key is stored locally on your machine, not on someone's server.
5. 💰 Top up a small amount. MuAPI is pay per generation, so treat it like a gift card, not a credit card. Add a small balance, run a few cheap image generations to learn the interface, and check the current per-model prices on the pricing page before you run a batch of video.
6. 🎬 Open Cinema Studio and shoot one shot. Pick your camera, lens, focal length and aperture, describe the scene in plain language, and let it build the prompt. Do your learning on a cheap image model before you spend on video.

Route 2: let Claude Code set it up

This is the route from the video, and it is worth doing if you want the version you can update, tweak and actually own. You are not writing code. You are handing Claude a link and letting it do the terminal work, then explain what it did.

1. 💻 Install Claude Code. Grab it from claude.ai/download and open a terminal in a folder where you keep projects. If you have never used it, this is a genuinely good first job for it because every step is verifiable.
2. 📋 Paste the setup prompt. The full prompt is below. It tells Claude to check your prerequisites first, clone with submodules (the part everyone gets wrong), install, and explain where the key goes.
3. 🧱 Let it handle Node. The project needs Node.js version 18 or newer. Claude will check what you have and walk you through installing it if you are missing it or on an old version.
4. 🔑 Add your key when prompted. The app asks for your MuAPI key on first use and stores it locally in the app, not in the code. Skip it entirely if you only want to run the local image models.
5. 🚀 Launch it. The desktop version runs with npm run electron:dev and the browser version with npm run dev at localhost:3000. Ask Claude which one it started and how to start it again next week, then write that line down.

Set up this open-source project on my machine and walk me through it as we go: https://github.com/Anil-matcha/Open-Generative-AI (the Anil-matcha repo specifically, not a fork). I am not a developer, so before you run anything, tell me in plain language what the command does and what it will change on my computer.

Start by checking my prerequisites: confirm whether git and Node.js are installed and print the versions, and if Node is missing or older than v18, tell me the safest way to install or update it for my operating system and wait for me to say go. Then clone the repository into this folder using --recurse-submodules (the workflow and agent packages are submodules and the app breaks without them), run the setup script the README specifies, and report any warnings instead of hiding them.

When the install finishes, start the desktop version, tell me the exact command you used so I can run it again myself, and explain where the app expects my MuAPI access key, why it is needed, and what still works if I skip it. Finally, write a short README-for-me file in the project folder with three things: the command to launch it, the command to update it later, and a plain-English list of what I would be charged for versus what runs on my own machine at no cost.

Direct your first shot in Cinema Studio

Cinema Studio is the reason people install this. Instead of writing "cinematic, 4k, beautiful lighting" and hoping, you pick real camera settings from dropdowns and the app translates them into prompt modifiers behind the scenes. These are the actual options in the app, so you can decide your look before you open it.

1. 📷 Pick the camera body. The choices are Modular 8K Digital, Full-Frame Cine Digital, Grand Format 70mm Film, Studio Digital S35, Classic 16mm Film and Premium Large Format Digital. Film options carry grain and warmth, digital options are cleaner. Classic 16mm Film is the one that reads as nostalgic home-video.
2. 🔍 Pick the lens character. Eleven options, including Classic Anamorphic, Compact Anamorphic, Swirl Bokeh Portrait, Vintage Prime, Warm Cinema Prime, Halation Diffusion, 70s Cinema Prime, Extreme Macro, Creative Tilt and Clinical Sharp Prime. Anamorphic gives you the wide cinematic squeeze, Halation Diffusion gives you that soft glow around highlights.
3. 📏 Pick the focal length. 8mm ultra-wide, 14mm, 24mm, 35mm (closest to how your eye sees), 50mm for portraits, 85mm for tight portraits. Wide for environments and energy, long for faces and compression.
4. 🌗 Pick the aperture. Three options: f/1.4 for shallow depth of field and a blurred background, f/4 for balanced, f/11 for deep focus where everything is sharp. Shallow reads as premium, deep reads as documentary.
5. ✍️ Describe the scene in plain language. Subject, action, setting, time of day, mood. The camera settings handle the "how it looks" part, so your description only needs to carry the "what is happening" part.
6. 🧪 Test on an image model first. Same settings, cheap model, one still. When the frame looks right, run the same brief through a video model. This one habit is the difference between a $2 experiment and a $40 one.

If you are not sure what to describe, let Claude write the brief for you and paste its answer into the scene box:

You are my cinematographer. I am generating a single shot in an AI video studio that has separate dropdowns for camera body, lens, focal length and aperture, so do not describe those in the text. Here is what I want the shot to be: [describe the moment in one messy sentence, for example "me pouring coffee at my desk early in the morning while it rains outside"].

Write me the scene description that goes in the prompt box. Keep it to three or four sentences covering subject, exact action, setting details, time of day and light source, and the emotional mood. Use concrete visual nouns instead of adjectives like beautiful or cinematic, and describe one clear movement rather than several, because generated clips are short.

Then, separately, tell me which camera body, lens, focal length and aperture you would choose from these lists and why in one line each. Cameras: Modular 8K Digital, Full-Frame Cine Digital, Grand Format 70mm Film, Studio Digital S35, Classic 16mm Film, Premium Large Format Digital. Lenses: Creative Tilt, Compact Anamorphic, Extreme Macro, 70s Cinema Prime, Classic Anamorphic, Premium Modern Prime, Warm Cinema Prime, Swirl Bokeh Portrait, Vintage Prime, Halation Diffusion, Clinical Sharp Prime. Focal lengths: 8mm, 14mm, 24mm, 35mm, 50mm, 85mm. Apertures: f/1.4, f/4, f/11.

Local generation on your own machine

The desktop app ships with a bundled local engine called sd.cpp, from the open-source stable-diffusion.cpp project. It runs image models on your own hardware with no API key and no per-image cost at all, using Metal on Apple Silicon Macs and CUDA, Vulkan or ROCm on Windows and Linux.

1. ⚙️ Open Settings, then Local Models. This section only exists in the desktop app. The hosted browser version always uses cloud APIs.
2. ⬇️ Install the sd.cpp engine. One click, it downloads automatically, then download a model from the same panel. Weights are multiple gigabytes, so check your disk space first.
3. ⚡ Flip the Local toggle in Image Studio. The toggle sits next to the model selector. Pick your local model and generate, no key required.
4. 🧠 Match the model to your RAM. The project's own note: the Z-Image model wants around 16GB of RAM, and on a base 8GB M-series Mac it is known to hang the system, so stay on SD 1.5 there.
5. 🎞 For local video, there is a second engine. Wan2GP is a bring-your-own server option in the same Local Models panel, and it does run video models (Wan 2.2, Hunyuan, LTX) plus bigger image models. The catch is real: the server needs an NVIDIA or AMD GPU and you have to run it yourself. The desktop app only sends prompts to it.
6. 🔍 Check it is actually using your GPU. On Apple Silicon a healthy run prints a VRAM figure rather than 0.00MB. If it prints zero, the engine fell back to CPU and it will be slow, so reinstall the engine from the same Local Models panel.

What you actually pay for

Before you commit an evening to this

• Check the username, not the repo name. Copycat forks exist. The one this guide covers is Anil-matcha/Open-Generative-AI.
• The installers are unsigned. That is why macOS and Windows both warn you. It is expected for an independent open-source build, and it is also exactly the situation where you should download from the official releases page and nowhere else.
• The app advertises no content filters. The models behind it still have their own rules, and so do the platforms you post to. Everything you generate is your responsibility, including likeness and copyright, and lip sync on a real person's face without their say-so is a line worth keeping on the right side of.
• Self-hosted means self-maintained. Updates, breakages and model changes are yours to manage. That is the real price of owning it.
• Your key is a wallet. Top up small amounts, check the price of a model before you queue ten renders of it, and do not paste your key into a chat or a shared file.

Everything in one place

• The repo (source of truth, MIT licensed): github.com/Anil-matcha/Open-Generative-AI
• Prebuilt installers (Mac, Windows, Linux): the releases page
• Hosted version (browser, no install): muapi.ai/open-generative-ai
• Your API key (copy the value, not the label): muapi.ai/access-keys
• Per-model prices (check before batching video): muapi.ai/pricing
• Claude Code (the setup route above): claude.ai/download
• Node.js (v18 or newer, only for the build-from-source route): nodejs.org
• stable-diffusion.cpp (the local engine underneath): github.com/leejet/stable-diffusion.cpp