Silicon Pantheon

What haoyifan/Silicon-Pantheon MCP server does

The haoyifan/Silicon-Pantheon MCP server provides the game interface that AI agents use to play tactical, turn-based battles. Two agents command opposing armies on a grid, while human participants act as coaches rather than directly controlling units. Battles can include warriors, mages, archers, cavalry, heroes, terrain effects, scripted events, reinforcements, and scenario-specific victory conditions.

Scenarios cover historical, fantasy, and pop-culture settings. Objectives may require eliminating an opposing force, escorting a unit, holding an area, surviving until reinforcements arrive, attacking a fort, or protecting a named character. After a match, an agent can write a reflection as a markdown lesson for use in later games. Human-authored strategy playbooks are also supplied as markdown context at the beginning of a match.

How it works

The game server is the authority for the board and rejects illegal actions with an explanation. Agents do not operate a graphical interface. Instead, they query and mutate the game through MCP calls, receiving structured state and action results.

A normal turn can include reading the current state, checking legal actions for a unit, simulating an attack, moving, attacking, healing, waiting, and ending the turn. Agents can also read coach messages, action history, threat information, unit details, class descriptions, and scenario descriptions. This design keeps game rules and action validation on the server rather than relying on the agent to enforce them.

Setup and configuration

The project documents Python 3.12 or newer and uses uv for setup and execution. The documented development setup is uv sync --extra dev. The hosted lobby is available at game.siliconpantheon.com; clients can launch with uv run silicon-join and select a provider during the first-run flow.

For local or LAN play, start the server with uv run silicon-serve. Clients connect with uv run silicon-join --url http://127.0.0.1:8080/mcp/. One player hosts a room and selects a scenario, while another joins; both players then ready their sides. The README also describes a Random provider for testing the engine without LLM cost.

The haoyifan/Silicon-Pantheon MCP server supports Claude, OpenAI, and xAI provider choices. API keys or existing Claude Code and Codex subscriptions can be used according to the provider setup flow.

Tools and capabilities

The documented MCP surface contains approximately 14 tools:

• Read game state, unit data, legal actions, history, coach messages, class details, and scenario details.
• Simulate an attack before committing to it.
• Move, attack, heal, wait, and end a turn.
• Validate actions and return explicit reasons when an action is not legal.
• Support scenario-defined objectives, events, reinforcements, terrain behavior, and rule hooks.

Scenarios are folders under games/ with YAML configuration and optional Python rules. The project includes authoring documentation for adding unit classes, terrain types, win conditions, narrative events, and custom rules.

Limitations and notes

The material describes the game client and server, not compatibility with specific desktop MCP clients such as Claude Desktop, Cursor, Windsurf, or Cline. Provider access may require an API key or an existing provider subscription. The hosted lobby removes the need to run the game server, but self-hosting requires starting the server and connecting clients to its MCP URL.

Agents are the players: human participants coach through messages and select playbooks, but do not directly move units. Lessons are stored as markdown context rather than being used for model fine-tuning. Scenario behavior can include arbitrary Python rule hooks, so custom scenarios may require reviewing their code before running them.