MCP Code Mode

Prototype implementation for the Code Execution MCP Server with DSpy. The "Code Execution with MCP" architecture combines the strengths of Large Language Models at code generation with the Model Context Protocol for tool integration. This system enables an AI agent to write Python code that runs in an isolated sandbox while seamlessly calling external MCP tools.

Quick Start

1. Installation

Requires Python 3.11+ and Node.js 20+.

2. Configuration

Copy the example environment file and configure secrets:

Configure your MCP servers in mcp_servers.json:

3. Running the Server

Launch the Code Execution MCP server:

4. Verification

Verify your setup by running the debug executor script. This script simulates an MCP client, connects to the server, and runs a test task to ensure the agent and tools are working correctly.

Before running the script:

1. Configure the MCP servers you want to interact with in mcp_servers.json.

2. Define the specific task you want the agent to perform by editing the task variable in scripts/debug_executor.py.

Development Commands

Execution Environment & Guardrails

By default, the system uses a Local Python Executor (LocalPythonExecutor) which runs code in the same process as the server. This is necessary because the strict Pyodide sandbox has limitations with network I/O, preventing it from calling back to other MCP tools in some environments.

Guardrails

Even with the local executor, the system enforces policies before code execution:

• Limits: 8k characters / 400 lines max.

• Imports: Allowlist only (json, math, re, datetime, etc.).

• Tokens: Disallows potentially dangerous tokens (subprocess, exec, eval).

Violations return a POLICY_VIOLATION error.

Note: You can force the use of the Pyodide sandbox by setting MCP_EXECUTOR=pyodide, but this may break tool calls depending on your environment.

Architecture

Overview

Why Code Mode?

Traditional MCP implementations face critical challenges:

1. Context Window Bloat: Every tool definition consumes tokens, limiting scalability.

2. Token Cost: Multiple back-and-forth tool calls are expensive.

3. Latency: Sequential tool invocations create cumulative delays.

4. Composability: Complex workflows require many discrete steps.

Code Mode addresses these by leveraging what LLMs excel at: writing code. Rather than making multiple tool calls, the agent writes a Python script that orchestrates all necessary operations internally.

Core Components

1. The Executor Server (FastMCP) (src/mcp_code_mode/executor_server.py) The server exposes an execute_code tool backed by a Python executor (Local or Pyodide). Uses fastmcp to handle the MCP protocol and dspy for execution logic.

2. Configuration-Driven Discovery (mcp_servers.json) The system uses mcp_servers.json to explicitly configure which MCP servers to connect to. Loaded by src/mcp_code_mode/mcp_manager.py.

3. Tool Schema Formatting (src/mcp_code_mode/tool_formatter.py) Formats discovered MCP tools into readable documentation that gets passed to the code generation LLM, so it knows what tools exist.

4. Context Injection The formatted tool schemas are passed as an input field to the LLM. The LLM knows tool names, parameters, and usage examples before it writes the code.

Information Flow

Troubleshooting

Timeout Issues: If the interpreter times out, it may enter a bad state. Currently, the best fix is to restart the server or reconnect the client to get a fresh interpreter instance.

Missing Tools: Ensure mcp_servers.json paths are correct and that you have run npm install if using Node-based servers.

References

• DSpy Documentation

• Model Context Protocol

• FastMCP

mcp-code-mode