gamecode-mcp by navicore - MCP Server

a naive but fun exploration into the MCP ideas ... but replaced with my next effort - see https://github.com/navicore/gamecode-mcp2

GameCode MCP Server

A Rust-based MCP (Model Context Protocol) server that provides dynamic CLI tool integration for Claude through a simple YAML configuration.

Caution: This is early-days for MCP and this is likely an irresponsibly naive implementation - use accordingly!!

Philosophy: Thoughtful Tool Design

LLMs famously don't do well with huge context windows even when they technically support them - so massive tool counts can erode performance.

Also, analogous to how you should design GraphQL schemas around personas and use cases rather than mirroring all your database schemas, you should thoughtfully select and configure your MCP tools. This is why we use YAML configuration instead of automatically consuming OpenAPI specs or DB connection strings for code generation.

Features

Dynamic CLI Tool Integration

The server loads all tools from a tools.yaml configuration file. This means:

No hardcoded tools - Everything is configurable
Add any CLI tool that returns JSON
Built-in tools included - Basic arithmetic and file operations come pre-configured
Hot reload - Just restart Claude Desktop after editing tools.yaml

Example Built-in Tools (from tools.yaml)

add - Add two numbers together
multiply - Multiply two numbers
list_files - List files in a directory

All tools are defined in your tools.yaml file - remove what you don't need, add your own!

Installation

Build and install the server:
```
cargo install --path .
```
This will install the gamecode-mcp binary to your Cargo bin directory (usually ~/.cargo/bin/).

Make sure your Cargo bin directory is in your PATH:

echo $PATH | grep -q "$HOME/.cargo/bin" || echo 'export PATH="$HOME/.cargo/bin:$PATH"' >> ~/.zshrc

Integration with Claude Desktop

Configure Claude Desktop using the provided script:

./scripts/configure_claude_code.sh

Or manually add to your MCP configuration:

{
  "mcpServers": {
    "gamecode": {
      "command": "gamecode-mcp",
      "args": [],
      "env": {}
    }
  }
}

Restart Claude Desktop to load the MCP server.

Configuration

The server looks for tools.yaml in these locations (in order):

Path specified in $GAMECODE_TOOLS_FILE environment variable
./tools.yaml (current directory - for project-specific tools)
~/.config/gamecode-mcp/tools.yaml (user defaults)

To get started:

# Create config directory
mkdir -p ~/.config/gamecode-mcp

# Copy the example configuration with all built-in tools and documentation
cp tools.yaml.example ~/.config/gamecode-mcp/tools.yaml

# Edit to customize and add your own tools
edit ~/.config/gamecode-mcp/tools.yaml

Adding Your Own CLI Tools

Edit your tools.yaml file to add new tools:

tools:
  - name: my_tool
    description: Description of what the tool does
    command: /path/to/your/tool
    args:
      - name: input
        description: Input parameter
        required: true
        type: string
        cli_flag: "--input"
    example_output:
      status: "success"
      data: "example output"

Then restart Claude Desktop to load the new tools.

See the examples/ directory for:

Complete tool configuration examples
Input/output mapping documentation
Real-world tool definitions

Testing

Once integrated, you can test the tools in Claude by asking:

"Use the gamecode run tool to execute add with a=5 and b=3"
"Use the gamecode run tool to execute multiply with a=4 and b=7"
"Use the gamecode run tool to execute list_files with path='.'"
"Use the gamecode list_tools to see all available tools"

Or more naturally:

"List all available gamecode tools"
"Add 5 and 3 using gamecode"
"Show files in the current directory using gamecode"

Development

For development, you can run the server directly:

cargo run

Audit Logging

The server supports optional audit logging of all tool invocations:

# Enable audit logging when configuring the MCP server
claude mcp add gamecode -s local ~/.cargo/bin/gamecode-mcp -- --audit-log ~/.config/gamecode-mcp/audit/

Audit logs are written to daily rotating files in the specified directory:

audit-2024-01-25.jsonl - One file per day
audit-2024-01-26.jsonl - Automatic rotation at midnight

The audit log records:

Timestamp - When the tool was invoked (UTC)
Tool name - Which tool was called
User - System username (from $USER environment variable)
Hostname - Device identifier for correlation with other logs

Example audit entries:

{"timestamp":"2024-01-25T10:30:00Z","tool_name":"add","user":"jsmith","hostname":"jsmith-macbook.local"}
{"timestamp":"2024-01-25T10:30:01Z","tool_name":"list_files","user":"jsmith","hostname":"lab-desktop-03"}

Security Considerations

Important: For security and privacy reasons, the audit log intentionally does NOT record:

Tool input parameters
Tool output/results
Any sensitive data

This design ensures that sensitive information (passwords, API keys, personal data, etc.) passed to tools is never persisted in audit logs. Organizations should correlate these audit records with other access control and monitoring systems to maintain full visibility while preserving security.

Troubleshooting

Server not appearing in Claude: Make sure the binary is in your PATH and Claude Desktop has been restarted.
No tools available: The server will show an error message with instructions if no tools.yaml is found. Check that your file is in one of the expected locations.
Tools not working: Check Claude Desktop logs (run with --mcp-debug flag) for error messages. Verify your CLI tools return valid JSON.
Configuration not loading: Ensure tools.yaml is valid YAML. You can test with: cat tools.yaml | yq '.'

License

MIT License - see LICENSE file for details