Pawball Arena

Inspiration

We wanted to explore what browser-native AI control can feel like when it is safe, legible, and fun. Pawball Arena turns that idea into a local arcade football match where an AI coach can understand the game semantically without gaining unrestricted control of the simulation.

What it does

Pawball Arena is a 3v3 cat football game with human play, local utility AI, and a WebMCP sideline coach. The coach can read the live match state, adjust formations and team style, call short tactical plays, assign roles, and issue bounded semantic orders such as marking an opponent or supporting a ball carrier.

The agent never sends raw key presses, coordinates, physics values, or score changes. Local AI remains responsible for movement, collision avoidance, timing, aim, stamina, and game legality.

How we built it

We built a browser-based local match simulation with six original players, each with a role, signature skill, stamina, and live AI intent. A tactical layer translates coach directives into bounded utility-AI effects.

WebMCP exposes 23 site tools for:

Reading match, pitch, player, event, and order state
Starting, pausing, and restarting matches
Setting formations, style, roles, marking, and skill policy
Calling short tactical plays
Issuing and tracking semantic player orders

The UI includes a visible coach opt-in, control-mode states, and a causality trace that shows tool input, validation, local effects, and order outcomes.

Challenges we ran into

The main challenge was preserving player agency while making AI coaching meaningful. Direct control is powerful but unsafe for a browser agent, so we had to define a semantic action layer that is expressive without exposing raw movement or physics mutation.

We also had to handle changing match state: possession changes, goal resets, expired directives, player control handoffs, and commands becoming illegal after they are issued. Clear rejection reasons and order lifecycle status were essential.

Accomplishments that we're proud of

A playable local 3v3 football experience with distinct teams and personalities
WebMCP tools that map intent to safe, local AI behavior
Visible consent gates for agent-assisted control
Real-time causality tracing instead of hidden automation
Orders that can be queued, executed, rejected, cancelled, expire, or end on a goal reset
A coaching interface that works alongside direct human play

What we learned

AI control works best when it operates at the same conceptual level as a human coach: “mark Cocoa,” “press high,” or “support the carrier,” rather than manipulating inputs directly.

We also learned that observability is part of trust. Showing why an order was accepted, rejected, or interrupted makes agent behavior understandable and debuggable.

What's next for Pawball Arena

Persistent match history, tactical replay, and coaching analytics
More formations, player traits, skills, and match scenarios
Better multi-turn coaching plans that adapt to score and possession
Expanded spectator and full-auto modes
Team customization, tournaments, and shareable match recaps
Additional WebMCP tools for strategic analysis while preserving the same consent and safety model