Documentation

SonexAI is a fully autonomous deep learning agent that perceives, reasons about, and acts within Valorant in real time. It is written in modern C++ with a CUDA inference backend and a policy trained through large-scale imitation learning and reinforcement learning against self-play.

This documentation covers installation, configuration, the runtime architecture, and the public API surface exposed by the client. It is intended for researchers and technical users who want to understand how the system operates end to end.

Getting SonexAI running takes four steps:

1. Create a folder anywhere on your Desktop.
2. Place the downloaded SonexAI executable inside it.
3. Launch the executable (running as administrator is recommended).
4. Complete the in-app setup wizard and enter your access key.

The client will verify your key, download the model checkpoint, and warm up the inference engine before the first match.

• Windows 10 / 11 (64-bit)
• Intel i5 8th gen / Ryzen 5 or newer CPU
• NVIDIA GTX 1660 or better with CUDA 12.x
• 8 GB RAM minimum, 16 GB recommended
• 2 GB free storage and the Visual C++ Redistributable

Runtime behavior is controlled through a config.toml file that sits next to the executable. Key parameters include the reaction profile, humanization level, and comms verbosity.

[agent]
reaction_profile = "competitive"   # instant | competitive | human | fatigued
humanization      = 0.7            # 0.0 (robotic) - 1.0 (very human)
target_fps        = 144

[comms]
enabled   = true
frequency = "standard"             # minimal | standard | verbose
personality = "concise"

Changes are hot-reloaded between rounds; no restart is required.

Each frame flows through a fixed pipeline with an end-to-end budget under 10ms: capture, preprocess, perceive, decide, and act. The perception stack produces a structured game-state tensor that the policy network consumes to output an action distribution.

For a full breakdown of every stage, network, and latency measurement, see the architecture reference.

The client exposes a local HTTP endpoint on 127.0.0.1:8787 for telemetry and control. All routes require the session token printed at startup.

GET  /v1/status        -> engine + model status
GET  /v1/telemetry     -> live decision traces
POST /v1/profile       -> switch reaction profile
POST /v1/pause         -> pause / resume the agent

Why does Windows flag the file? The client injects a model into a live process, which heuristic antivirus tools misread as malware. It is a false positive — see the download page for how to allow it.

Do I need an access key? Yes. A one-time access key is required to unlock the forum and run the client.