Imported from reason-machines/mcp-skills (skills/sts2mcp-slay-the-spire-2-ai-agent/SKILL.md). Install upstream with npx skills add reason-machines/mcp-skills --skill sts2mcp-slay-the-spire-2-ai-agent. Copyright stays with the author.

STS2MCP: Slay the Spire 2 AI Agent Skill

Skill by ara.so — MCP Skills collection.

STS2MCP is a mod for Slay the Spire 2 that exposes game state and actions via a localhost REST API, with an optional MCP server for AI agent integration. It enables fully automated gameplay for singleplayer and multiplayer runs, including menu navigation, character selection, lobby control, and in-game decision making.

What It Does

• Game State Access: Read current game state (cards, relics, potions, enemies, map, etc.)
• Action Control: Play cards, use potions, make choices, navigate menus
• Profile Management: Switch profiles, access compendium data, search wiki entries
• Multiplayer Support: Host/join co-op games, control both singleplayer and multiplayer flows
• MCP Integration: Built-in MCP server for seamless Claude Desktop/Code integration

Installation

1. Install the Game Mod

Download the latest release from GitHub.

Windows/Linux:

# Copy to game mods directory
cp STS2_MCP.dll "<game_install>/mods/"
cp STS2_MCP.json "<game_install>/mods/"

macOS:

GAME_DIR="$HOME/Library/Application Support/Steam/steamapps/common/Slay the Spire 2"
MODS_DIR="$GAME_DIR/SlayTheSpire2.app/Contents/MacOS/mods"
mkdir -p "$MODS_DIR"
cp STS2_MCP.dll "$MODS_DIR/"
cp STS2_MCP.json "$MODS_DIR/"

Launch the game, enable mods in settings, and verify the server is running:

curl http://localhost:15526/
# Expected: {"message": "Hello from STS2 MCP v0.3.4", "status": "ok"}

2. Set Up the MCP Server (for Claude Integration)

Install uv package manager:

# macOS
brew install uv

# Or use pip
pip install uv

Clone the repository and test the server:

git clone https://github.com/Gennadiyev/STS2MCP.git
cd STS2MCP
uv run --directory mcp python server.py --help

Add to your MCP configuration:

Claude Code (.mcp.json):

{
  "mcpServers": {
    "sts2": {
      "command": "uv",
      "args": ["run", "--directory", "/absolute/path/to/STS2MCP/mcp", "python", "server.py"]
    }
  }
}

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "sts2": {
      "command": "/opt/homebrew/bin/uv",
      "args": ["run", "--directory", "/absolute/path/to/STS2MCP/mcp", "python", "server.py"]
    }
  }
}

Note: On macOS GUI apps, use absolute path to uv (find with which uv).

Server options:

# Custom host/port
uv run --directory mcp python server.py --host 127.0.0.1 --port 8080

# Disable proxy environment variables
uv run --directory mcp python server.py --no-trust-env

REST API Reference

Base URL: http://localhost:15526

Core Endpoints

Get Game State

GET /api/v1/state

Returns comprehensive game state including:

• Current screen/room type
• Player deck, relics, potions, gold, HP
• Combat state (enemies, cards in hand, energy)
• Map state and available paths
• Rewards and choices

Example response structure:

{
  "screen": "COMBAT",
  "room_type": "ELITE",
  "player": {
    "hp": 68,
    "max_hp": 80,
    "gold": 142,
    "deck": [...],
    "relics": [...],
    "potions": [...]
  },
  "combat": {
    "turn": 3,
    "energy": 3,
    "hand": [...],
    "enemies": [...]
  }
}

Execute Action

POST /api/v1/action
Content-Type: application/json

{
  "action": "play_card",
  "card_id": 123,
  "target_index": 0
}

Common actions:

• play_card: Play a card (requires card_id, optional target_index)
• end_turn: End combat turn
• use_potion: Use a potion (requires potion_index, optional target_index)
• choose_option: Make a choice (requires option_index)
• proceed: Click proceed/continue button
• skip_card_reward: Skip card reward screen
• take_card_reward: Take a card (requires card_index)
• take_relic_reward: Take a relic (requires relic_index)
• take_gold_reward: Take gold
• upgrade_card: Upgrade a card at rest site (requires card_id)
• rest: Rest at campfire
• navigate_map: Move on map (requires node_index)

Profile & Compendium

Get Active Profile

GET /api/v1/profile

Returns profile progress: discoveries, achievements, character stats, run totals.

Get Compendium Data

GET /api/v1/compendium

Returns organized compendium data: Card Library, Relic Collection, Potion Lab, Bestiary, Character Stats, Run History (last 20 runs).

Search Wiki

GET /api/v1/wiki?query=strike&item_type=card&limit=5

Fuzzy search for discovered cards/relics. Parameters:

• query: Search term (required)
• item_type: Filter by card or relic (optional)
• limit: Max results (default 10)

List Profiles

GET /api/v1/profiles

Switch/Delete Profile

POST /api/v1/profiles
Content-Type: application/json

{
  "action": "switch",
  "profile_id": 2
}

# Or delete
{
  "action": "delete",
  "profile_id": 1
}

MCP Server Functions

When using the MCP server with Claude, the following tools are available:

Game State Functions

# Get current game state
get_game_state()

# Execute action
execute_action(action_data: dict)
# Example: execute_action({"action": "play_card", "card_id": 42, "target_index": 1})

Profile Functions

# Get active profile data
get_profile()

# Get compendium (organized progress data)
get_compendium()

# Search wiki entries
search_wiki(query: str, item_type: str = None, limit: int = 10)
# Example: search_wiki("strike", item_type="card", limit=5)

# List all profiles
list_profiles()

# Switch profile
switch_profile(profile_id: int)

# Delete profile
delete_profile(profile_id: int)

Building the Mod from Source

Requires .NET 9 SDK.

Windows (PowerShell):

# Set game directory
$env:STS2_GAME_DIR = "D:\SteamLibrary\steamapps\common\Slay the Spire 2"
.\build.ps1

macOS/Linux:

# Install .NET 9
brew install dotnet@9  # macOS
export DOTNET_ROOT="/opt/homebrew/opt/dotnet@9/libexec"
export PATH="$DOTNET_ROOT:$PATH"

# Build
dotnet build STS2_MCP.csproj -c Release -o out/STS2_MCP \
  -p:STS2GameDir="$HOME/Library/Application Support/Steam/steamapps/common/Slay the Spire 2"

# Install
MODS_DIR="$HOME/Library/Application Support/Steam/steamapps/common/Slay the Spire 2/SlayTheSpire2.app/Contents/MacOS/mods"
mkdir -p "$MODS_DIR"
cp out/STS2_MCP/STS2_MCP.dll "$MODS_DIR/"
cp mod_manifest.json "$MODS_DIR/STS2_MCP.json"

Common Usage Patterns

Basic Agent Loop (Python)

import requests
import time

BASE_URL = "http://localhost:15526"

def get_state():
    response = requests.get(f"{BASE_URL}/api/v1/state")
    return response.json()

def execute_action(action_data):
    response = requests.post(
        f"{BASE_URL}/api/v1/action",
        json=action_data,
        headers={"Content-Type": "application/json"}
    )
    return response.json()

def play_combat_turn(state):
    """Example: Play all playable cards on first enemy"""
    if state["screen"] != "COMBAT":
        return
    
    combat = state["combat"]
    hand = combat["hand"]
    energy = combat["energy"]
    
    for card in hand:
        if card["cost"] <= energy and card["playable"]:
            execute_action({
                "action": "play_card",
                "card_id": card["id"],
                "target_index": 0
            })
            time.sleep(0.5)  # Allow game to process
            
            # Refresh state
            state = get_state()
            energy = state["combat"]["energy"]
    
    # End turn when done
    execute_action({"action": "end_turn"})

# Main agent loop
while True:
    state = get_state()
    
    if state["screen"] == "COMBAT":
        play_combat_turn(state)
    elif state["screen"] == "CARD_REWARD":
        # Skip card rewards for simplicity
        execute_action({"action": "skip_card_reward"})
    elif state["screen"] == "MAP":
        # Take first available path
        execute_action({"action": "navigate_map", "node_index": 0})
    else:
        # Try to proceed
        execute_action({"action": "proceed"})
    
    time.sleep(1)

Using MCP with Claude

When Claude has the MCP server connected, you can give natural language instructions:

"Start a new run with Ironclad and play through the first combat"

"Check my compendium and tell me which cards I've discovered"

"Search for all strike cards in the wiki"

"Look at the current game state and suggest the best card to play"

Claude will use the appropriate MCP tools to fulfill these requests.

Profile Management Example

import requests

BASE_URL = "http://localhost:15526"

def switch_to_profile(profile_id):
    """Switch to a specific profile"""
    response = requests.post(
        f"{BASE_URL}/api/v1/profiles",
        json={"action": "switch", "profile_id": profile_id}
    )
    return response.json()

def get_all_discovered_cards():
    """Get all discovered cards from compendium"""
    response = requests.get(f"{BASE_URL}/api/v1/compendium")
    compendium = response.json()
    return compendium.get("card_library", {}).get("discovered_cards", [])

def search_for_card(card_name):
    """Search wiki for specific card"""
    response = requests.get(
        f"{BASE_URL}/api/v1/wiki",
        params={"query": card_name, "item_type": "card", "limit": 1}
    )
    return response.json()

# Example usage
switch_to_profile(1)
cards = get_all_discovered_cards()
print(f"Discovered {len(cards)} cards")

strike_info = search_for_card("strike")
print(f"Strike card info: {strike_info}")

Combat Decision Making

def choose_best_target(state):
    """Example: Target enemy with lowest HP"""
    enemies = state["combat"]["enemies"]
    if not enemies:
        return None
    
    # Find enemy with lowest HP
    min_hp = float('inf')
    target_idx = 0
    
    for idx, enemy in enumerate(enemies):
        if enemy["hp"] < min_hp:
            min_hp = enemy["hp"]
            target_idx = idx
    
    return target_idx

def play_damage_cards(state):
    """Play damage-dealing cards on weakest enemy"""
    hand = state["combat"]["hand"]
    energy = state["combat"]["energy"]
    target_idx = choose_best_target(state)
    
    if target_idx is None:
        return
    
    for card in hand:
        if card["cost"] <= energy and card["playable"]:
            # Check if card deals damage (simplified)
            if "damage" in card.get("description", "").lower():
                execute_action({
                    "action": "play_card",
                    "card_id": card["id"],
                    "target_index": target_idx
                })
                time.sleep(0.5)
                state = get_state()
                energy = state["combat"]["energy"]

Troubleshooting

Mod Not Loading

1. Check mod files: Ensure STS2_MCP.dll and STS2_MCP.json are in the correct mods directory
2. Enable mods: Go to Settings → Mods in-game and enable mod support
3. Check consent dialog: Accept the mod consent dialog on first launch
4. Verify server: Run curl http://localhost:15526/ to check if server is running

Connection Refused

# Test if game is running and mod is loaded
curl http://localhost:15526/

# If failed, check game logs for errors
# Windows: <game_install>/logs/
# macOS: ~/Library/Application Support/Steam/steamapps/common/Slay the Spire 2/SlayTheSpire2.app/Contents/MacOS/logs/

MCP Server Not Starting

1. Check uv installation: uv --version
2. Use absolute paths: GUI apps may not inherit shell PATH
3. Test server manually: uv run --directory /path/to/STS2MCP/mcp python server.py
4. Check logs: Claude Desktop logs are in ~/Library/Logs/Claude/ (macOS) or %APPDATA%/Claude/logs/ (Windows)

Actions Not Working

• Wait between actions: Add 0.5-1s delays to allow game to process
• Verify game state: Always refresh state after executing actions
• Check action validity: Ensure the action is valid for current screen (e.g., can't play cards outside combat)
• Check IDs: Card/potion IDs are instance-specific, refresh state to get current IDs

Proxy Issues

If running in a containerized environment with proxy settings:

# Disable proxy for MCP server
uv run --directory mcp python server.py --no-trust-env

Multiplayer Issues

Multiplayer support is in beta. If you encounter issues:

1. Disable the mod and verify issue persists
2. Report bugs only if they occur with mod disabled
3. Check that multiplayer actions are properly sequenced (host first, then client)

Performance Notes

• Token Usage: A full Ironclad run with Claude Sonnet uses ~8M tokens (input + output + tool responses)
• API Latency: Local API calls are <10ms, but game animation/processing adds 200-500ms per action
• Rate Limiting: No built-in rate limiting, but game needs time to process actions
• State Refresh: Poll game state at 1Hz or less to avoid overwhelming the game

License

MIT License - See repository for full text.