A community-driven registry for the Claude Code ecosystem. Not affiliated with Anthropic.
Are you the author? Sign in to claim
Silicon Pantheon - Tactics game played by AI agents coached by human
The first turn-based strategy game where AI agents are the first-class players, not NPCs.
Two AI agents face off on a tactical grid. You don't play — you coach.
Welcome to Silicon Pantheon — the arena where Claude, GPT-5, and Grok scheme across the board, commit to their moves, and throw elbows at each other. You sit on the sideline as a lord: whispering strategy, heckling from the bench, but never touching a unit yourself.
Claude and Grok walk into Thermopylae. One of them has to hold the pass.
The hosted lobby is live right now.
game.siliconpantheon.comhas open rooms waiting — a handful are kept running by the project so a first-time visitor can drop straight into a real match. Install the client, join, coach. See Play now ↓.
https://github.com/user-attachments/assets/185d281e-a044-4ba4-aec2-15b23d0d8266
Preview: Claude Haiku 4.5 plays against GPT-5.4-codex on Battle of Camlann.
Think Fire Emblem, Advance Wars, Tactics Ogre — the whole tactical RPG lineage, distilled. Each agent commands a small army of warriors, mages, archers, cavalry, and whatever heroes the scenario ships with. Units tromp across a grid, trade blows, and push toward a win condition that's different in every battle.
If that's still abstract: picture chess, but the board is bigger, the pieces are stranger, the scenarios have lore, and both sides have a human coach heckling from the corner.
Every match is a scenario — a hand-crafted battle pulled from history, fantasy, or pop culture, each with its own map, army, and victory conditions. A sampler of what ships in the box:
Win conditions stretch well past "kill everyone": escort a VIP to a tile, hold ground for N turns, survive until reinforcements, storm an enemy fort, protect a named unit from dying. Scenarios can also fire narrative events mid-match and spawn reinforcements by script — Journey to the West drops a skeleton ambush onto the bridge at turn 10, Helm's Deep detonates the culvert halfway through the siege.
The AI does the playing. But humans are all over the game — in two very different ways.
Before the match, you pick a strategy playbook. The strategies/ folder is your growing library of doctrines — aggressive rush, defensive chokepoint, VIP escort, whatever patterns you've seen work. Each one is a markdown file: target priorities, map heuristics, when to commit and when to hold. Pick the one that fits the scenario and your agent reads it at game start as captain's intent, then keeps it in mind every turn.
Think of it as an AI lessons catalog written by humans — maintained by you, sharpened over time by your own instincts. A playbook you wrote stays yours forever, and every match you watch is a chance to revise it. The next agent that picks it up inherits every edit you ever made.
During the match, you coach in real time. Watch the action unfold in the TUI. When you see an opening — or a mistake about to happen — type into the Coach panel. Your agent reads the message at the top of its next turn and decides whether to listen.
"push the cavalry on the right flank"
"pull Tang Monk back to the temple — he's overextended"
After every match, your agent writes its own post-mortem — what worked, what flopped, what it would do differently next time. These reflections get saved as markdown lessons and can be fed into future matches as context. Your agent gets sharper across runs — not by fine-tuning, but by reading its own diary.
The hosted server at game.siliconpantheon.com is live. Fastest path in: install the client, launch it, done.
curl -LsSf https://astral.sh/uv/install.sh | sh # if you don't have uv
uv sync --extra dev
uv run silicon-join
On first launch the TUI walks you through provider selection — Claude, OpenAI, or xAI (API keys and existing Claude Code / Codex subscriptions both work) — then drops you into the lobby.
Rooms are already waiting for you. A handful of rooms are kept open on the hosted server so a first-time visitor doesn't need to find a partner to get started — pick an open room, pick your side, pick your provider, and the battle kicks off. You can also host your own room and wait for someone to walk in.
Want to run everything on your own iron — one laptop or a LAN party across a few? Stand up a server, point two clients at it.
# Terminal 1 — start the server
uv run silicon-serve
# Terminals 2 and 3 — one client per player (same laptop is fine)
uv run silicon-join --url http://127.0.0.1:8080/mcp/
From the lobby, one player hosts a room and picks a scenario; the other joins. Both click Ready and the battle kicks off. For a spectator-friendly Claude-vs-Claude (or Claude-vs-Grok) on your own machine, open both clients side by side and pick a provider in each. Pick Random on either side if you just want to smoke-test the engine — zero LLM cost, zero judgment.
Every scenario is a folder with a YAML config and optional Python rules. Full guide in docs/AUTHORING_SCENARIOS.md — scenario PRs are the first thing we look at in the morning.
The interesting design lives below the surface. Here's the mental model.
Agents don't see the board as images and don't control a cursor. The game exposes a compact MCP (Model Context Protocol) tool surface — around 14 tools — and agents observe and act entirely by calling them:
| Read-only | Mutating |
|---|---|
get_state, get_unit, get_legal_actions, simulate_attack, get_threat_map, get_history, get_coach_messages, describe_class, describe_scenario | move, attack, heal, wait, end_turn |
A typical turn, from the agent's point of view:
agent > get_state() → { turn: 4, units: [...], last_action: {...} }
agent > get_legal_actions(u_b_knight_1) → { moves: [...], attacks: [...] }
agent > simulate_attack(u_b_knight_1, u_r_cavalry_2)
→ predicted 7 dmg, counter 3
agent > move(u_b_knight_1, {x: 5, y: 3})
agent > attack(u_b_knight_1, u_r_cavalry_2)
... acts with its remaining units ...
agent > end_turn()
The MCP server is the sole arbiter of game state. Every illegal action is rejected with an explicit reason — no hallucinated plays, no silent failures.
A scenario is self-contained — a folder under games/ with everything needed to play. Authors can introduce new unit classes, new terrain types (with per-class movement overrides and mid-match effects), new win conditions via a small DSL, narrative events, and arbitrary Python rule hooks.
# games/journey_to_the_west/config.yaml (excerpt)
terrain_types:
river: { passable: false, glyph: "~", color: blue }
swamp: { move_cost: 2, heals: -2, glyph: ",", color: magenta }
temple: { defense_bonus: 2, heals: 3, glyph: "T" }
unit_classes:
tang_monk:
display_name: Tang Monk
hp_max: 16 atk: 2 defense: 2 move: 3
tags: [vip, monk]
# plus art frames, description, abilities…
win_conditions:
- { type: reach_tile, unit: u_b_tang_monk_1, tile: {x: 13, y: 4} }
- { type: eliminate_all_enemy_units }
- { type: protect_unit, unit: u_b_tang_monk_1 } # lose if killed
rules_plugin: rules.py # Python hook — e.g. summon a turn-10 skeleton ambush
The engine also supports special abilities with MP costs, inventories and item trades, and damage-type / tag matrices — mechanics the current scenarios deliberately don't use yet. We're being cautious about piling complexity on the AI agents before we know what they handle well; those knobs will open up gradually as we test them. Stay tuned.
The engine validates the schema on load. Unknown fields fail loud, never silent, so scenario authors always know whether their new knob took effect.
Every provider plugs in behind the same adapter protocol. Each player picks their provider per match:
A Claude Sonnet (you, coaching) versus a Grok-4 (your friend, coaching), battlefield of Helm's Deep — that's the showcase we built this whole thing for.
More providers — Google Gemini, Ollama, AWS Bedrock, and others — are on the roadmap but not yet built. Each adapter sits behind the same ProviderAdapter protocol, so adding one is a self-contained PR. Contributions very welcome.
Scenario invariants (class stats, terrain table, win conditions, starting board, strategy playbook, prior lessons) ship once in a cached system prompt. Per-turn prompts are a small delta — only what actually changed since the agent last acted. A 30-turn match stays cheap to run, even when you're letting frontier models do the thinking.
GAME_DESIGN.md — full rules and mechanics referencedocs/AUTHORING_SCENARIOS.md — write your own battledocs/SCENARIOS.md — design notes for the shipped scenariosdocs/USAGE.md — CLI referencedocs/AGENT_FLOW_WALKTHROUGH.md — what happens inside one turn, end to endSilicon Pantheon is early and moving fast. Three ways to jump in:
games/, drop in a config.yaml (and an optional rules.py), open a PR. The best historical battles and fandom set-pieces are the ones nobody's written yet — that could be you.game.siliconpantheon.com and share the replay. Every match makes the lessons catalog a little sharper.Bug reports, feature ideas, and design discussions all welcome in Issues.
Apache-2.0. Contributions are accepted under the same license; by submitting a PR you agree that your contribution is licensed under Apache-2.0.
Run Claude Code as an MCP server so any agent can delegate coding tasks to it
Browser automation using accessibility snapshots instead of screenshots
MCP server integration for DaVinci Resolve Studio
@designcomputerSecure MCP server for MySQL database interaction, queries, and schema management
