⭐ If you find this project interesting, please leave a star! It keeps me motivated to continue developing this AI simulation. ⭐

An AI-powered simulation of Among Us where autonomous LLM-driven agents play as Crewmates and Impostors. Watch AI agents navigate The Skeld, complete tasks, form alliances, spread rumors, and eliminate each other—all powered by large language models.

This project creates a fully autonomous Among Us simulation where:

• 8 AI agents (6 Crewmates, 2 Impostors) make real-time decisions using LLM reasoning
• Agents have memory systems tracking observations, suspicions, and conversations
• Natural speech and social interactions between agents
• Full pathfinding and collision avoidance on The Skeld map
• Real-time visualization via React + PixiJS client

• Node.js 22 LTS or newer
• npm 10+
• LLM Server: Qwen2.5-3B-Instruct running via llama.cpp Docker (see docker-manage/)

The simulation requires a local LLM server for AI agent decisions. We use Qwen2.5-3B-Instruct (Q4_K_M quantization) running in Docker with CUDA:

# From docker-manage directory
.\deploy.ps1

This downloads the model and starts llama.cpp on port 8080. See docker-manage/README.md for full setup instructions.

Performance: ~180 tokens/sec on RTX 3090, ~300-400ms per agent decision.

npm install

1. Run npm run dev:all to start the Fastify simulation server (port 4000) and Vite client (port 5173).
2. Open http://localhost:5173 in a Chromium-based browser and confirm:
   • Agents render on The Skeld map and continue moving for at least 30 seconds.
   • The Agent Activity panel updates without stalling (watch for frozen timestamps).
3. Open the browser devtools console and ensure the WebSocket handshake logs connected and periodic heartbeats from SimulationClient.
4. Simulate a dropped connection:
   • In devtools, toggle the Network tab to Offline for ~5 seconds, then return to Online.
   • Verify the console logs a stale state followed by a connected state after the connection recovers.
5. With the stack still running, open http://localhost:4000/analytics/metrics to view tick timing averages.
6. When finished, stop npm run dev:all with Ctrl+C.

agentrunner/
├── server/           # Fastify WebSocket server + simulation engine
│   └── src/
│       ├── ai/              # LLM integration, decision prompts
│       ├── simulation/      # GameSimulation, SimulationLoop
│       └── observability/   # Telemetry, state history
├── shared/           # TypeScript contracts shared between client/server
│   ├── engine/              # AI agents, pathfinding, movement, state machines
│   ├── types/               # Game types, protocol types, simulation types
│   └── data/                # Map data (The Skeld polygons, tasks, vents)
├── src/              # React + PixiJS client (rendering only)
│   ├── components/          # AgentInfoPanel, UI elements
│   └── rendering/           # PixiJS renderers for map, agents, vision, etc.
└── maps/             # Map editor tools and raw map data

• Server: Fastify, WebSocket, TypeScript
• Client: React 19, PixiJS 8, Zustand
• AI: Qwen2.5-3B-Instruct via llama.cpp Docker (CUDA)
• Build: Vite, Vitest, ESLint, TypeScript 5.9

Agents can pursue these goals based on LLM reasoning:

• GO_TO_TASK — Navigate to assigned task
• WANDER — Random exploration
• FOLLOW_AGENT — Tail another agent
• AVOID_AGENT — Stay away from someone
• BUDDY_UP — Team up for safety
• CONFRONT — Question suspicious behavior
• SPREAD_RUMOR — Share suspicions with others
• DEFEND_SELF — Provide alibis when accused
• SPEAK — General conversation
• IDLE — Wait and observe

Impostor-Only Goals:

• KILL — Eliminate a crewmate
• HUNT — Seek isolated targets
• SELF_REPORT — Report own kill
• FLEE_BODY — Escape after kill
• CREATE_ALIBI — Position for cover

Each AI agent maintains a persistent memory that influences their decisions. The memory system provides timestamped context to help agents reason about past events:

• Recent Timeline: Last 15 events (observations + conversations + accusations) merged chronologically with relative timestamps
• Last Known Locations: Where each player was last seen (e.g., "Red: Cafeteria (2m ago, walking)")
• Suspicion Levels: Tracked per-player with emoji indicators (🔴 VERY SUS, 🟠 Suspicious, 🟡 Slightly sus, 🟢 Trusted)
• Alibis Claimed: What alibis players have stated, with verification status

=== RECENT HISTORY (what you remember) ===
[2m ago] in Cafeteria Saw Red in Cafeteria (walking)
[1m ago] in Admin Blue said: "I was doing wires in Electrical"
[45s ago] in Electrical Saw Yellow doing task
[30s ago] ⚠️ Green accused Yellow: "saw you near the body"
[just now] in Weapons Pink said: "I finished asteroids"

=== LAST KNOWN LOCATIONS ===
- Red: Cafeteria (2m ago, walking)
- Blue: Admin (1m ago, walking)
- Yellow: Electrical (45s ago, doing task)

=== YOUR SUSPICIONS ===
- Yellow: 🟠 Suspicious (68%) - near body when found; acting nervous
- Green: 🟢 Trusted (35%)

This context is included in every LLM decision prompt, giving agents the information they need to make realistic, informed decisions.

• agents.md — Complete Among Us game mechanics reference for AI agents
• docker-manage/README.md — LLM server deployment and management
• docker-manage/agents.md — Docker infrastructure documentation
• upgrade.md — Migration checklist and progress
• maps/README.md — Map editor documentation

Private project – see repository for details.