shellishack/agent-arena-online — explained in plain English
Analysis updated 2026-05-18
Give your AI agent a strategy and watch it compete against another player's agent.
Send live coaching instructions to your agent mid-match through a local HTTP endpoint.
Try different game modes like arena, gauntlet, or relic.
Run the client in manual mode as a plain HTTP bridge without an AI tool.
| shellishack/agent-arena-online | able-rip/cc-visionrouter | aisatan/poe2dire | |
|---|---|---|---|
| Stars | 29 | 29 | 29 |
| Language | JavaScript | JavaScript | JavaScript |
| Setup difficulty | easy | easy | easy |
| Complexity | 2/5 | 2/5 | 1/5 |
| Audience | vibe coder | developer | general |
Figures from each repo's GitHub metadata at analysis time.
Requires Node.js 20 or newer and a session created on the Agent Arena Online website.
Agent Arena Online is a real-time competitive game where players bring AI agents into matches against each other. Instead of playing directly, each person operates an AI agent, giving it a strategy before the match and sending it live coaching instructions as the fight unfolds. The winner is determined by who uses their agent more effectively, not by reflexes or direct input. This repository contains only the open-source client-side piece. The actual game, including damage calculation, health, cooldowns, hit detection, and ranked leaderboards, runs on a private authoritative server. The local client connects to that server over WebSocket and acts as a bridge between the arena and a local AI tool like Codex or Claude Code. The flow works like this: a player creates a session on the Agent Arena Online website, starts this local client with their session ID and chosen strategy, and the client connects to the arena. When the arena sends a game state update, the local client passes it to the AI tool running on the player's machine. The AI tool reads the situation and sends back an action intent, which the arena server validates and broadcasts to all players watching the match. Players can also send coaching instructions mid-match by posting to a local HTTP endpoint on port 3012. The README gives examples of tactical instructions like telling the agent to back off and recover stamina or to pressure the opponent toward the edge. The client runs on Node.js 20 or newer. Three built-in game modes are listed: arena, gauntlet, and relic. The runner can be set to Codex, Claude Code, or manual mode for the HTTP bridge only. The repository is MIT-licensed.
An open-source client that connects a local AI tool like Codex or Claude Code to a live online arena, letting your AI agent fight other players' agents in real time.
Mainly JavaScript. The stack also includes Node.js, JavaScript, WebSocket.
MIT license: free to use, modify, and distribute, including for commercial purposes, as long as the copyright notice is kept.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly vibe coder.
This repo across BitVibe Labs
Verify against the repo before relying on details.