Open sourcing Game Summoner, my prompt to game site

tldr: Open sourced my prompt to game suite: https://github.com/ndamiano/ai-agent-test, it's MIT licensed, and this runs well on my 5090 with 64gb ram, but any model that can handle tool calls can manage. Edit: I can't believe I forgot to share a game... This is one shot with qwen 3.8 flash next! https://gamesummoner.com/g/siWy5VIMxF-H Hey everyone, I recently launched https://gamesummoner.com. You may have also seen that google just released https://playground.google. I cannot compete with that, and honestly, I wanted to figure out how I could give back to the people here who, probably unknowingly, helped me get from idea to implementation. This isn't a nice clean repo for you to trivially run with something like python run.py, as it is tailored to my specific setup on digital ocean, and using runpod and aws as my hosted GPUs. That said, there is a script called scripts/local_gpu.py that will get you most of the way to running this locally. It manages the worker boxes, spinning up instances of inference (I use ninfer with Qwen 3.8 27B locally and a modified sglang https://github.com/ndamiano/sglang-rtxpro6000 with Qwen 3.8 Flash-Next on a rented rtx 6000 in prod, and comfyui with a buncha models), but this can be modified by your agent to launch however you need. The architecture is pretty straightforward. Anytime a request comes in, it throws it into a queue (sql, I am cheap, and it works just as well at this scale as something like kafka), a worker long polls for work, picks it up, and returns the value. This requires there to be workers, and so I built a simple autoscaler, it walks up the cost ladder from aws / runpod to try to get the cheapest GPU available (I probably should add more sources, but eh, that's work on the least interesting part.) And for how the actual generation goes, I've done a ton of iterations (you can see many of them in https://github.com/ndamiano/maestro-labs, as I said.. several iterations on the name), and settled on creating a design doc with a team of agents. The first agent creates a high level design, second and third in parallel are visual and engineering, fourth is an integrator that puts them all together as the "holy grail" of the design. Once we've got the design, in it goes to the same model, with a new prompt, that is, effectively, build the game described in the design. We give it access to tools that let it test the game, take screenshots, etc. and wait for it to call done. Once it's finished, we validate the build and give it a quick "play", where the model looks at a photo, tries some input, and sees what happens. We return any exceptions and inputs that do nothing (the model has notoriously been AWFUL at "is this good"...), and once there are none, we say "complete" and return to the user. There are a couple other repos that are necessary: https://github.com/ndamiano/gamesummoner-workers https://github.com/ndamiano/gamesummoner-images (I told you, the name went through some iterations...) All said and done, I'm releasing this with an MIT license. This is a full, scalable, deployable website that generates games. I made sure all of the models used are well licensed, and so should probably not be an issue if you want to stand it up. There's quite a bit of setup, but like, you could get this up and running in a couple days with an agent. If you do and somehow make a few million, I'm currently unemployed, so I'd love a job lol.