The Intent Horizon

An Asimov-inspired WebMCP RTS strategy game about why agentic instruction matters: bits take years to reach Alpha Centauri and Earth’s view is already old, so Daneel needs intent, not clicks.

Comment

Inspiration

WebMCP has a story problem, demos are mostly for storefronts, operating tools, reduce computer use context burn, but they don't demonstrate what its useful for beyond the base ideas as of now. With this game I thought, I would help make a demo for WebMCP as a game, to show what agentic things can achieve and help overcome even such long intent horizons and constraints such as even the speed of light.

I read Foundation when I was young. I have been binge-watching the show recently, and it brought back the part of those stories that always stuck with me: people making decisions for futures they will never live to see.

That feeling led to The Intent Horizon.

I wanted to build a city builder where distance makes ordinary control break down. You are on Earth, managing a colony at Alpha Centauri. Every image of the colony is 4.37 years old by the time you see it. If you respond with a mouse command, that command takes another 4.37 years to arrive.

A normal instruction can be perfectly reasonable and still fail. By the time it reaches the colony, the tile may be flooded, the workers may be elsewhere, or the problem may have changed completely.

So Earth sends Daneel something more useful: a goal, constraints, and a little room to act.

Earth sends intent: goals, constraints, and preferences. Daneel uses WebMCP tools to interpret that intent against the colony’s current reality.

Keep food reserves above 24 months.
Keep power above 20%.
Protect the wetlands.
Prefer expanding the existing agricultural district.

Daneel receives that message beside the colony as it actually is, then has to decide what to do with it.

The game is not a Foundation adaptation. Its influence is more personal than that. I wanted to make a small game about long horizons, delegated power, and the uneasy question underneath both: if you cannot control the future directly, what should you ask for?

What it does

The Intent Horizon is a city builder where Earth and a local AI steward play different roles.

Earth sees an old reconstruction of the colony and decides what it values. Daneel sees the current colony and uses WebMCP tools to inspect it, build infrastructure, assign robots, manage resources, send reports, and request more authority when a decision goes beyond the original instruction.

The radio channel is limited in both directions. Reports take years to return. Earth does not get a secret live feed. A local survey, construction job, or crisis stays unknown until telemetry makes the trip home.

There are three missions:

  1. The First Light: Build enough life support and redundant power to survive a planned outage.
  2. The Meaning of Enough: Keep food and power reserves healthy without sacrificing protected wetlands.
  3. The Right to Decide: Export iridium while deciding how much local freedom is safe around a fragile habitat.

How we built it

I built the game with React, Vite, and Three.js. It is fully client-side and stores the simulation in versioned localStorage.

The same deterministic simulation powers the UI and the WebMCP tools. There is no game backend, database, custom MCP server, or game-owned model API.

The colony tracks delayed packets, radio bit budgets, construction jobs, robot labor, roads, power networks, food and water reserves, seeded hazards, authority requests, and mission outcomes. If Daneel builds something, it still needs materials, a connected site, available labor, and time.

The WebMCP tools are game actions, not mouse clicks. Daneel can inspect the colony and resource network, read delivered mail, construct buildings, build roads, register limited policies, send reports, request authorization, yield control, and wait for the next useful event.

When native WebMCP is available, the game uses it directly and makes its availability visible to the player. If WebMCP is not there, the user can still play and control the colony manually and give instructions via mouse clicks like any other RTS game.

Challenges we ran into

The difficult part was making the two perspectives feel honest.

If Earth can see the live colony, the distance is cosmetic. If Daneel can act without limits, the player has no real role. If reports appear immediately, the game loses its tension.

I had to keep drawing the line between what Earth knows, what Daneel knows, and what neither can safely assume. Earth sees only received telemetry. Daneel acts locally, but is still limited by resources, roads, workers, power, hazards, and the authority Earth has actually granted.

Making the wait playable took work too. The game shows packet cost, arrival dates, transmission windows, Earth-visible events, captured dates, and receipt dates. The delay needs to be something the player can plan around, not a loading screen.

Native WebMCP support also varies by client, model, account, and browser. The game checks for it and says so plainly when it is unavailable. It does not pretend that a scripted helper is a live agent.

Accomplishments that we're proud of

I am happiest with the fact that WebMCP is part of the game’s premise, rather than an extra feature.

The human chooses the values. Daneel works inside the present reality. Both have limits.

The missions gradually move away from simple execution. First, the player asks whether Daneel can carry out a local plan. Then they learn whether they can communicate a resilient policy. Finally, they have to decide what authority they are willing to give away when they will not be able to correct the outcome in time.

What we learned

The project changed how I think about agent tools.

Giving an agent access to every button is less interesting than giving it a small set of meaningful actions with clear costs and limits. inspect_colony and construct_building are useful because they describe what is happening in the world and let the system validate the action. A generic click tool would not.

It also made the human role clearer. The player is not there to out-compute the agent. They are there to decide what should not be optimized away.

"Protect the wetlands" and "prefer reversible interventions" are not ordinary resource targets. They are choices about what kind of colony Earth is trying to create.

What's next for The Intent Horizon

I want to make the unknown parts of the colony matter more.

Surveys should reveal terrain and discoveries that change the player’s understanding of the situation. Daneel’s reports should explain the tradeoffs behind a decision, not just list what was built. I also want to add more shared codebooks and doctrines, where a short phrase carries meaning because Earth and Daneel have earned that vocabulary together.

The question I want to keep exploring is simple:

When the future is far away and you only get a few messages, how do you say what you mean?

Built With

Share this project:

Updates