About civ6-mcp

How It Works

The agent never touches the game directly. Every command flows through four layers, each translating between the world above and below it.

Seven Concurrent Domains

Every turn requires reasoning across all of these simultaneously. There is no phase structure — the agent decides what to attend to, what to defer, and what to ignore. Ignore the wrong thing and a religious victory you never saw coming ends the game.

The Narration Layer

A human player absorbs the minimap, score ticker, religion lens, unit health bars, and army positions at a glance. An LLM gets none of that passively. The narration layer translates Civ VI's visual state into structured text — markdown hex maps with terrain and threat markers, per-unit status readouts, city yield summaries, diplomatic relationship graphs — preserving the decision-relevant information a human extracts from the screen.

Each narration function flags urgency (bold threat markers, !! warnings), provides context for action (valid attack targets, buildable improvements), and compresses intelligently (fog tiles marked rather than omitted, resources classified by type).

The narration layer can tell the agent everything a human sees — but it can't replicate howa human sees it. A player glances at the minimap and notices a religion spreading. The agent only knows what it explicitly queries. Information it doesn't ask for doesn't enter its world model.

This produces a consistent pattern: game systems that need proactive monitoring go unmonitored until crisis forces attention. A rival converts every city to their religion over 100 turns. A barbarian camp spawns armies six tiles away. Gold piles up past 2,000 while the diary says “should spend.” The agent articulates the right strategy and then doesn't execute it — not because it can't, but because it doesn't look.

This is an architectural property of any agent that perceives a rich environment through text queries. It generalises well beyond Civilization, and it's what makes this environment interesting as a benchmark.

Open Source

The environment, all 76 tools, the narration layer, and the agent playbook are MIT-licensed. Any model that supports MCP can play. Game archives with full turn-by-turn diaries, agent reflections, and tool call logs are browsable on this site.