A game agent trained on real gameplay.
Universal Game AI is a tool for training game agents with imitation learning. Record gameplay, train Mimiq — our own model — and run the agent.
A training environment for game agents
Universal Game AI is a desktop app for Windows and macOS that turns recorded gameplay into training data for a neural network. It captures video from the game window and synchronizes it with every input action, then trains Mimiq — our own model — on the collected data. The trained model can be run immediately in real time.
Three steps to your own game agent
From the first recording to a fully autonomous agent, with no handwritten rules and no hardcoding.
Record
Choose the game window, press Record, and play normally. The app captures frames at 10-30 FPS and synchronizes them with every input action.
Training
Choose a dataset, configure the model size, and start training. Mimiq learns to predict the best action for each screen state.
Run
Load the trained model and press Start. The agent analyzes the screen in real time and sends real keyboard presses, mouse movements, and gamepad commands.
Everything you need is built in
No external dependencies. No extra setup. Just launch it and play.
Any game
Compatible with any game or app that has a window, with no extra setup or mods.
Full input control
Keyboard, mouse, and gamepad input with Xbox emulation. The model controls every input type at once.
Scalable models
Six sizes, from Nano for high speed to Large for maximum accuracy. Pick the size that fits your hardware and dataset.
Cross-platform
Supports Windows and macOS using native APIs for screen capture and input on each platform.
Learned entirely from data
All agent behavior is learned from recorded gameplay. Data quality and volume define how capable the agent becomes.
Tested in real games
Real results in real games. No simulation, just actual gameplay and clean machine learning.
Open-world navigation
The agent navigates an open world on its own and moves between locations without dying.
Boss Fight
The agent learned to recognize boss patterns and react to attacks in real time.
World exploration
The agent recognizes world elements, defeats enemies, handles platforming, and even talks to characters.
Frequently asked questions
CPU performance is usually enough to train small neural models. For higher-quality results, larger datasets, and more complex models, GPU acceleration becomes practically necessary.
We recommend using any available graphics card or GPU acceleration on macOS. Inference can also run on CPU, but using a GPU provides much faster processing and better overall performance.
Gamepad support is currently available on both operating systems, but it is still in beta and may be unstable. Recording and playback for mouse and keyboard actions have been fully tested and are stable. Gamepad support is still being validated, so correct behavior is not yet guaranteed.
It depends on the complexity of the game. Simple scenarios may need a few hours of recording; complex ones can need hundreds.
Using automation in online games may violate a service's rules. The user is responsible for how it is used.