MCP Code Execution Server

What it does

This server implements a 'discovery-first' architecture for MCP. Instead of preloading hundreds of tool schemas into the LLM's context (which can cost thousands of tokens per request), it exposes a single run_python tool. The LLM writes Python code to dynamically discover, call, and compose other MCP tools on demand.

Tools

• run_python: The primary entry point. Executes Python code in a secure, rootless sandbox. Within the code, the LLM can use helpers like discovered_servers(), query_tool_docs(), and search_tool_docs() to find and execute other MCP tools.

Installation

Add the following to your claude_desktop_config.json:

{
  "mcpServers": {
    "mcp-server-code-execution-mode": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/elusznik/mcp-server-code-execution-mode",
        "mcp-server-code-execution-mode",
        "run"
      ],
      "env": {
        "MCP_BRIDGE_RUNTIME": "podman"
      }
    }
  }
}

Supported hosts

• Claude Desktop
• Cursor
• VS Code Copilot
• Windsurf