FORGE — Human-Agent World Laboratory

Inspiration

Game teams use disconnected tools to inspect world state, reproduce progression failures, rebalance encounters, approve risky changes, validate fixes, and recover from regressions. General-purpose browser agents add friction when they must infer domain meaning from pixels or manipulate fragile UI controls.

FORGE explores a better model: one shared, typed control plane where humans and browser agents can understand and safely operate on the same interactive world.

What it does

FORGE is a human-governed live-operations and QA laboratory demonstrated through Ashen Reach, a playable medieval RPG world.

A WebMCP agent can:

Inspect locations, characters, quests, gates, items, encounters, and player state Analyze the real quest graph and discover progression blockers Run deterministic, seeded combat and reachability simulations Propose narrowly scoped world repairs Execute approved changes through the same service layer as the human interface Validate the result, inspect provenance, and roll back to a prior checkpoint

The primary demonstration contains a real circular dependency: the key required to open the Ashen Sanctum is spawned inside that locked sanctum. The agent finds the defect from structured world state, proposes relocating the key, and waits for a person to approve it. A regression run then proves the repaired quest route is reachable. Rollback restores the original checkpoint and makes the defect reproducible again.

Why WebMCP is essential

FORGE exposes 30 browser-native capabilities through document.modelContext.registerTool(). These tools span world inspection, player actions, quests, encounters, seeded simulation, QA, repair, and governance.

WebMCP gives the agent typed domain capabilities instead of forcing it to guess from pixels or button labels. Closed schemas validate nested input, stable identifiers are checked against the current world, and mutation tools clearly advertise their approval boundaries.

The human interface and every WebMCP tool call the same typed application service. That shared boundary performs validation, permission checks, proposal routing, checkpoint creation, execution, audit recording, and activity publication. WebMCP is the product interface—not a chatbot placed on top of the product.

A better human-agent experience

FORGE deliberately separates capability from authority:

OBSERVE: agents may inspect and simulate, but cannot mutate the world PROPOSE: agents may prepare consequential changes, but a person must approve them AUTONOMOUS: bounded changes may execute immediately only after a human explicitly opts in

Every reversible mutation creates a checkpoint. Activity and audit views distinguish native WebMCP agents, human actions, approvals, validation runs, demo helpers, and registry examples. A person can inspect parameters, modify them, reject a proposal, approve it, or restore a previous world state without erasing governance history.

This creates a workflow that was previously fragmented across debug tools, scripts, dashboards, and manual playtests:

Structured inspection → bounded proposal → human approval → measured execution → deterministic validation → audit and rollback

How we built it

FORGE uses React and TypeScript with a client-first, deterministic world model. The WebMCP registry and human UI share a single executeForgeTool application service. The QA layer performs reference validation, quest-graph validation, progression reachability analysis, and seeded combat simulation.

The project includes:

30 registered WebMCP tools with schemas and lifecycle cleanup Recursive runtime validation for nested tool input Human approval, rejection, parameter editing, checkpoints, audit, rollback, and deterministic reset Seeded combat simulation and measured difficulty calibration Responsive desktop and narrow ChatGPT-style layouts 19 service-level tests and a Playwright end-to-end matrix Public-origin verification through ChatGPT’s in-app browser Security headers, open-source licensing, CI, and zero known production dependency vulnerabilities

How we used AI and Codex

Codex helped evolve the concept into a production-ready implementation. It reviewed the shared-service architecture, traced authorization and persistence edge cases, strengthened nested validation, replaced misleading metrics with measured methodology, designed difficulty calibration, improved the accessible judge flow, added automated testing and CI, diagnosed rollback and hydration issues, and verified the public deployment.

Human judgment remained responsible for product positioning, public access, approval boundaries, evidence review, and final release decisions.

Try it

Open the live app in ChatGPT’s in-app browser or Chrome with WebMCP enabled. Open the WebMCP inspector and confirm all 30 capabilities. Ask the agent to find the most serious progression blocker and propose the narrowest reversible repair without applying it. Review and approve the proposal as the human. Run the regression and confirm the deadlock disappears. Inspect the audit trail, roll back, and retest to reproduce the original defect.

No credentials are required.

Links

Live app: https://forge-human-agent-world-laboratory.themfpretty.chatgpt.site Source: https://github.com/jpretty01/forge-webmcp Demo video: https://youtu.be/wIQB_eFURDI

Known limitations

WebMCP is an evolving Chromium preview API and is not available in every browser. World state is currently isolated in browser storage rather than a shared multi-user backend. Combat simulation is deterministic and abstract rather than a frame-by-frame game-engine replay.