How to use Codex to build an asymmetric WebMCP game
Codex built a two-level WebMCP puzzle with unequal human and agent information, dynamic tools, and reconnection; its creator retained art and playtesting.
devpost.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Build an asymmetric PixiJS co-op puzzle where the human operates machinery and Codex reads bounded telemetry through WebMCP. Expose tools only when systems become available, enforce information boundaries in game state, and support limited-resource diagnostics. Test dynamic registration, browser disconnection, checkpoints, and reconnection, leaving art selection and playability review to me.
Reach people exploring AI agents
See placements →What happened
GROUNDTRUTH became a two-level asymmetric puzzle game with a live Codex partner, state-dependent tools, checkpoints, and a reconnect flow.
Reported by Vishal TK — DevpostHow the agent did it
- Vishal used Codex to implement the TypeScript and PixiJS puzzle game and its WebMCP connection.
- Capabilities appeared and disappeared with restored systems and puzzle progress, preserving unequal player information.
- The developer playtested connection failures, added reconnection, and selected and corrected generated assets.
Where a person stepped in
Vishal directed mechanics, playtested the game, rejected unsuitable assets, and repeatedly corrected visuals; the human player retained physical puzzle actions.
Vishal TK built GROUNDTRUTH as a two-player puzzle in which the second player is a Codex agent. The human explores a damaged ship and operates machinery, while the agent reads telemetry, searches records, and runs limited diagnostics through WebMCP. Neither side receives all the information.
The developer built and refined the game inside one Codex thread. In TypeScript and PixiJS, tool availability became part of the mechanics: restoring a system exposed its telemetry, while unopened or offline sections remained unavailable. State and tools enforced those boundaries instead of relying on the agent to remember a long instruction.
Development exposed concrete integration problems. Tool-registration limits required capabilities to be added and removed as puzzles progressed. Unexpected browser closures disconnected the agent, so the game gained a reconnect path. Generated visual assets also needed extensive human selection and correction for noise, transparency, scale, and readability.
The reported result was a complete two-level experience with checkpoints, reconnection, win and failure sequences, and a participating agent. The practical lesson is that narrow live tools can define meaningful cooperation. Visual consistency and playability still came from repeated human judgment, and the account does not establish reliability across every browser or agent harness.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Devpost · Vishal TK
Frequently asked questions
Does the agent see everything the player sees?
No. Human visual access and agent telemetry are deliberately different.
Why do tools change during play?
Offline or unreached systems should not expose their information.
How did the project handle lost connections?
The developer added a reconnect flow instead of requiring a restart.
More details
About this account
Vishal used Codex to implement the TypeScript and PixiJS puzzle game and its WebMCP connection. Capabilities appeared and disappeared with restored systems and puzzle progress, preserving unequal player information.
Opened original account and sanvishal/ground-truth repository. Source’s enormous token counter includes cached tokens and is omitted. Runtime cooperation and implementation are both evidenced; harness stability and manual asset corrections are preserved. Suggested prompt written by Agents In Use; no original task prompt was published in the verified account.
Credit: Vishal TK — Devpost
Source published 3 Sept 2026
Source checked 14 Oct 2026
Tools and topics
Codex · WebMCP · PixiJS · TypeScript · image generation
co-op puzzles · tool availability · reconnection
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.