AI Trainer Simulator

A 3D retro pixel-art browser game where you are an IT trainer at a quirky office in Warsaw. Walk around, talk to coworkers, deliver training, run standups, debug client scripts, and try not to go bankrupt. Full vision in docs/PRD.md; the shared roadmap is in docs/plans/game-roadmap.md.

"Make this the best simulator business retro game in the history, a real game, not just a demo." — Lucas

WebMCP: play alongside an AI agent

This game exposes its own gameplay to a browser-resident AI agent through WebMCP. The agent does not screenshot the page and guess at pixels - it calls real tools that drive the same code paths the human player's input drives.

The agent is a player, not an administrator. It can join as a visible robot coworker, look around, walk to named people and places (by name, never by coordinate), step and turn like a human on WASD, and speak. It cannot give itself money, set flags, teleport, move your camera, or answer your dialogue choices for you - those tools do not exist.

The capability worth looking at is agent-authored dialogue: when you talk to the robot, the game hands the agent the conversation context, and the agent writes that character's spoken line and the reply options you are offered. Those lines were never written by this game's author. It is an LLM-driven NPC whose inference runs in your browser, so the game ships no API key and costs nothing per player.

Trying it

1. Open the game in a browser with WebMCP available - ChatGPT's browser, or Chrome with chrome://flags/#enable-webmcp-testing set to Enabled and relaunched.
2. Verify it is live before you start: the title screen prints Agent play ready - N WebMCP tools live. If it says agent play is unavailable, the browser has no model-context surface and the game will still play normally, just without the agent. The browser console carries the same result under [webmcp].
3. Ask your agent to play along - for example, "read the game's instructions, then join as a coworker called Rusty who is a sarcastic QA engineer, walk over to Bartek, and say hello." The get_instructions tool hands it the whole protocol, so it does not have to guess.
4. Walk up to the robot and click it to start a conversation - or let the agent open one itself with start_conversation. Conversations work in both directions.

How it is built

• src/webmcp/tools.ts - the tool implementations and their validation.
• src/webmcp/agent-dialogue.ts - the conversation handshake, including the long-poll (wait_for_player_message) that lets the agent hear a reply the moment it happens. WebMCP has no way for a page to push to an agent, but tool calls are awaited, so holding one open is the closest thing to a notification - and it degrades into ordinary polling when it times out.
• src/engine/agent-companion.ts - the robot itself: pathing, walking, speaking, and the raw movement controls.
• src/webmcp/bridge.ts - registration with the browser. Probes document.modelContext, then navigator.modelContext, then the testing shim, because the specification is still moving and sources disagree on the namespace.
• docs/ADR/0008-webmcp-browser-bridge-and-agent-companion.md - why it is built this way, including why the agent companion deliberately sits outside the NPC scheduling system.
• docs/PRD-hackathon-webmcp.md - the requirements and acceptance criteria.

Registration never throws. A browser without WebMCP support reaches a fully playable game with no visible difference.

How to run the game

Live preview (RECOMMENDED — has HMR)

pnpm dev

Open http://localhost:5173/ in your browser.

Every time you save a .ts / .css file, the page reloads automatically (Vite HMR). This is the URL to use during development. The dev server is already running; you don't need to start it.

Production preview (static build on port 4173)

pnpm build          # one-off production build
pnpm build:watch    # rebuilds dist/ on every save
pnpm preview        # serves dist/ at http://localhost:4173/

pnpm preview is a static file server — it does NOT watch for source changes. To see changes on 4173, you must run pnpm build (or pnpm build:watch) first.

Common mistake (Lucas, 2026-08-29): testing on 4173 while the agent edits source on 5173 → "the fix didn't work!" — but it did, 4173 was just serving the old dist/. Use 5173 for live preview.

Run both side by side (live preview at 5173, production at 4173)

# Terminal 1: dev server with HMR (live preview)
pnpm dev

# Terminal 2: rebuild dist/ on every save
pnpm build:watch

# Terminal 3: serve dist/ at 4173 (optional, for "real" production build)
pnpm preview

Tests

pnpm typecheck   # TypeScript only, fast
pnpm test        # vitest, unit tests (68 tests as of 2026-08-29)
pnpm test:e2e    # Playwright e2e smoke

Project layout

src/             game source (TypeScript)
tests/unit/      vitest unit tests
tests/e2e/       Playwright end-to-end tests
docs/PRD.md      what we're building (with corrections log)
docs/ADR/        architecture decisions
docs/plans/      shared roadmap and code-bound plans (status stays in Beads)
AGENTS.md        agent rules (read this if you're a new agent)

Project status

The game has progressed beyond the original numbered phase snapshot, so status is intentionally not duplicated here. Current work and known issues live under Beads epic sacs-xtma; the product progression is in docs/plans/game-roadmap.md, and Lucas's complete append-only feedback record is in docs/LUCAS-FEEDBACK-INDEX.md.