CodexGame
github.com/dimillian/codexgame- Category
- Entertainment
- Rank
- No. 2054Tools index
- Pricing
- Open Source
- Type
- GAME
- Builder
- dimillian
- GitHub
- 15 stars
- Date
About
Dimillian's experiment: trap Codex agents in a game and play god — watch how AI agents behave in a constrained synthetic environment.
What it does
CodexGame runs a self-playing isometric RPG where one to four Codex-controlled characters act inside a simulated world. Agents receive limited world snapshots, then choose structured actions such as moving, gathering, crafting, fighting, talking, inspecting peers, and changing relationships. A browser interface shows the map, activity feed, scores, inventories, social state, and controls for pausing, resetting, messaging agents, and selecting models.
Why it's ranked here
This is a focused, inspectable agent experiment with more substance than a visual demo. The world is deterministic, actions are schema-validated, and survival, progression, and social behavior have explicit mechanics. The clear boundary between agent output and simulation state makes behavior easier to observe. Its value remains specialized: setup requires local development tools and a working Codex command-line installation.
What's good
The constrained action vocabulary turns vague model output into observable game decisions. Seeded world generation supports repeatable runs, while validation limits malformed turns and content edits. Agents can negotiate, form alliances, inspect inventories, loot opponents, craft items, place structures, and fight creatures. The interface exposes model choice, reasoning effort, latency, scores, relationships, inbox messages, and nearby entities, giving experiments useful context rather than only animated characters.
Tradeoffs
CodexGame is a private workspace package, not a published installable application. Running it requires Node.js 20 or newer, pnpm 10 or newer, and the Codex command-line tool. The documented endpoints bind to local loopback addresses, and no hosted deployment workflow appears in the supplied text. Scores and prompt guidance encode strong preferences for survival, gathering, crafting, social actions, and deliberate combat, so observed behavior reflects those designed incentives rather than an unconstrained agent.
How to use it well
Use CodexGame for hands-on study of structured agent decisions, repeatability across seeds, model comparisons, and experiments with cooperation or conflict. Start with identical agents and settings, change one model or effort level, then compare actions, latency, scores, and social outcomes. Mutable recipes, prefabs, biome rules, and spawn rules support controlled scenario changes. It does not replace a general agent evaluation suite, production orchestration platform, or hosted game service.
Technical notes+
The root package.json defines a private TypeScript ESM pnpm workspace with recursive build, test, lint, and typecheck scripts. apps/game-runtime supplies the authoritative runtime, Codex bridge, and WebSocket API; apps/game-client uses React, Phaser, and Vite; packages/protocol contains Zod transport contracts; packages/simulation implements seeded generation, snapshots, scoring, and state transitions. packages/protocol/src/messages.ts defines a versioned WebSocket protocol for sessions, snapshots, actions, feeds, builds, pause and reset controls. packages/protocol/src/build.ts and packages/protocol/src/actions.ts expose both Zod schemas and JSON Schema objects, but their JSON Schema representations use nullable all-field records rather than mirroring the discriminated Zod variants directly.
Observed
- Primary language
- TypeScript
- Packaging
- Private pnpm workspace using ECMAScript modules
- Install surface
- Source installation with pnpm; requires Node.js 20+, pnpm 10+, and the Codex CLI
- User interface
- Local React and Phaser web client served by Vite
- Runtime interface
- Versioned WebSocket API on a local runtime server
- Architecture
- Separate runtime, web client, shared protocol, deterministic simulation, and mutable content workspaces
Read from README.md, package.json, packages/protocol/src/index.ts, packages/protocol/src/build.ts, packages/simulation/src/rng.ts, packages/simulation/src/index.ts, packages/protocol/src/actions.ts, packages/simulation/src/types.ts, packages/simulation/src/world.ts, packages/protocol/src/messages.ts, packages/simulation/src/snapshot.ts, packages/simulation/src/simulation.ts, eslint.config.js, apps/game-client/vite.config.ts, apps/game-client/src/App.tsx.
What it can do
Create synthetic game environment for AI agents
Game parameters and constraints → Constrained virtual environment
Deploy Codex AI agents into game world
AI agent configurations → Active AI agents in environment
Monitor AI agent behavior and interactions
Agent activities and decisions → Behavioral data and interaction logs
Control environmental variables and constraints
Environment modification commands → Modified game conditions
Track agent decision-making processes
Agent reasoning and choices → Decision analysis reports
Generate behavioral analytics from agent interactions
Raw agent interaction data → Behavioral patterns and insights
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.