The tool really is free. Generating is not, and there is exactly one way to stay free

The repo is here. It is called Open Generative AI, MIT licensed and open source. 28,122 stars, 4,994 forks, last updated on the day I wrote this post. So it is not an abandoned project, it is moving fast. The desktop app is at v1.0.9 with ready made installers for Mac, Windows and Linux.

What is extra below: which path is actually free and where you start burning credits, how to set up the local path that does not even ask for an API key, verified install commands, the model that freezes 8 GB Macs, the step the guide skips when installing the Claude Code skills, and whose interest this repo actually serves.

This is the most important section of the post. The app is free, no argument there: you download it, install it, pay nothing, and the source is open. But the moment you generate something, there are two separate worlds. Is the model running on your machine or in the cloud? The answer decides the price.

The cloud side runs on MuAPI. I checked MuAPI's free endpoints: there are exactly five and none of them generate images or video. Image watermarking, video watermarking, a Gemini audio profile, character creation, and Suno voice cloning while it is in preview. Their own wording is clear: image, video and most audio models are pay per use. New accounts get a small trial credit allowance, but the amount is not published anywhere, so I am not quoting a number. You will see it in your own dashboard.

The app ships with an engine called sd.cpp. You download a model once, then generate unlimited images on your own machine. It does not ask for an API key and does not even need the internet. The guide covers this in a single line in its troubleshooting section, yet it is exactly what anyone who wants to stay free actually needs.

Video is a different story and it is worth correcting your expectations upfront. To generate video locally you have to run a separate server called Wan2GP, and that server is CUDA bound, meaning an NVIDIA or AMD card is required. There is no Apple Silicon path. If you are on a Mac you cannot generate video locally; you can only install it on a Windows or Linux box on your network and point the app at it. On top of that, the repo notes that wiring Video Studio to this engine is not finished yet and sits on the roadmap.

The install really is one click and two minutes. But both Mac and Windows throw a warning on first launch because the app is unsigned. Both are normal and both are a one time fix.

# Open the DMG, drag the app into Applications, then:
xattr -cr "/Applications/Open Generative AI.app"

# Then right click the app, choose Open, and click Open again in the dialog.
# If you would rather avoid Terminal: System Settings, Privacy and Security,
# scroll down and click "Open Anyway".

On Windows you get the SmartScreen screen. Click More info, then Run anyway. The app installs silently and drops a Start Menu shortcut. Since the source is open, this warning is not about the code itself but about the installer being unsigned.

This was the part that drew the most attention in the video, and it is real. But knowing what it actually is makes your life easier: this is not a camera simulation, it is an interface that writes better prompts for you. The camera, lens and aperture you pick turn into phrases appended to the prompt. That is how the repo describes it too.

Knowing this helps, because you can build the same logic by hand in other tools. The options: six camera bodies, eleven lens characters, six focal lengths from 8mm to 85mm, and three apertures. f/1.4 melts the background, f/11 keeps everything sharp, and 35mm is the one closest to the human eye.

The guide covers this in two sentences and trips exactly here. The skills themselves are real, the repo is live and MIT licensed. But the install is not what the guide says, and the difference matters, because without the skipped step the skills install fine and then fail to generate anything.

# 1. The command line tool
npm install -g muapi-cli

# 2. The API key (the step the guide skips)
muapi auth configure
# You get the key from the muapi.ai dashboard

# 3. Install the skills into your agent
npx skills add SamurAIGPT/Generative-Media-Skills --all

# If you only want it in Claude Code
npx skills add SamurAIGPT/Generative-Media-Skills --all -a claude-code

I am not writing this as a takedown, the tool is real and it works. But using it knowingly is a different thing. There is a MuAPI badge at the very top of the repo, and the second section of the README is a straight sales page: resell this under your own brand at your own prices, starting at $49 a month. So this project is not a neutral community effort, it is MuAPI's customer acquisition channel.

Knowing this explains two things. First, why the word free is everywhere and the credits part sits at the bottom. Second, why the no install browser version always runs on the cloud: that version is the storefront. Being open source does not change any of this, it just leaves you the option to stay local. That option is real, as long as you know where it is.