LeagueBot

An automated League of Legends bot. A C++ Windows client plays matches end-to-end using an ONNX vision model and a kernel-level virtual HID driver for input; Python tooling records gameplay and trains the models.

Status: personal / research project. Running game-automation software against Riot's client violates the League of Legends Terms of Service and will get accounts banned. Use at your own risk, on accounts you are willing to lose.

---

Repository layout

ai/              C++ bot (Visual Studio, x64)
  bot/           capture → ONNX inference → play loop
  client/        League client launcher + screen-coord helpers
  cloud/         GitHub-backed account rotation & config fetch
  hid/           Driver install + user-mode HID I/O
  lcu/           League Client Update API glue
  resource/      Embedded assets (ONNX models, vhidev driver, onnxruntime.dll)
gen/             Small codegen helper
recorder/        Python screen+input recorder for building datasets
train/           PyTorch training + ONNX export
test/            Standalone HID / mouse smoke test
third_party/     Vendored onnxruntime SDK (headers + libs)

Components

ai/ — bot runtime (C++)

Single UAC-elevated Windows executable. On launch it:

1. Checks whether the vhidev virtual-HID driver is installed; if not, extracts the embedded driver package from its own RCDATA and installs it (persists across reboots).
2. Opens the HID interface and takes over mouse/keyboard output.
3. Pulls account credentials and runtime config from a private GitHub repo via a fine-grained PAT (XOR-obfuscated at build time; see ai/cloud/token_secret.h.example).
4. Launches the League client, logs in, queues, and hands off to the play loop (ai/bot/play.cpp), which runs the ONNX model embedded as ai/resource/{blue,red}.onnx.

recorder/ — dataset capture (Python)

recorder.py captures screen frames with mss plus mouse/keyboard events with pynput, writing side-by-side recordings into recorder/recordings/{blue,red}/.

train/ — model training (Python / PyTorch)

Trains the per-side vision + action model and exports to ONNX for the C++ runtime to consume. Entry points: train.py (training), export.py (PyTorch → ONNX), play_test.py (sanity-check an exported model).

Build

C++ bot

Requirements:

• Visual Studio 2022 (v143 toolset), Windows 10/11 SDK
• Target platform: x64, configurations: Debug / Release

Steps:

1. cp ai/cloud/token_secret.h.example ai/cloud/token_secret.h and paste a GitHub Fine-grained PAT with read access to your private config repo.
2. Open ai.sln and build ai | x64 | Release.
3. Run the produced ai.exe as administrator (the manifest already requests elevation).

Python tooling

cd recorder && pip install -r requirements.txt
cd ../train && pip install -r requirements.txt

Python ≥ 3.10 recommended. For GPU training, install a CUDA-matching torch build before running pip install -r train/requirements.txt.

Notes on secrets

• ai/cloud/token_secret.h is gitignored and must never be committed.
• The PAT is XOR-scrambled into the exe to resist strings scraping — this is deterrence, not security. Rotate the PAT if the binary leaks.

License

No license granted. All rights reserved. Code is published for reference.