VRChat Agent

Low-frequency intent + high-frequency instinct loop VRChat agent.
The agent observes screen/audio, plans with LLM at controlled frequency, and executes actions via OSC/local input.

Features

• Idle instinct loop (continuous micro-actions, no per-frame LLM dependency)
• Intent gating (LLM called on scene/heard change or TTL)
• Startup preflight checks (OSC / window / audio / API)
• Runtime presets (quiet, active)
• Optional memory retrieval (data/memory.jsonl)
• F11 extra speak / F12 stop hotkeys

Requirements

• Windows 10/11
• Python 3.11+
• VRChat with OSC enabled

Install

python -m pip install -r requirements.txt

Configure

1. Copy config template:

Copy-Item config\config.example.toml config\config.toml

2. Copy env template:

Copy-Item .env.example .env

3. Fill .env:

SILICONFLOW_API_KEY=your_real_key

Run

python main.py --help
python main.py --preset quiet
python main.py --preset active

--preflight-only: TODO (not implemented as standalone flag yet).
Use scripts/run_preflight.ps1 or scripts/run_preflight.bat for startup preflight checks.

VRChat Setup (short)

• Enable OSC in VRChat settings
• Keep OSC host/port aligned with config (127.0.0.1:9000 by default)

Notes

• This is a WIP project.
• PRs are welcome.
• Please keep contributions incremental; no architecture rewrite-only PRs.