MCP Code Execution Server

Install

Works in

• Claude Code
• Claude Desktop
• Cursor
• OpenCode
• Windsurf
• VS Code
• SourceCraft
• Koda
• Coddy

Manual install

{
  "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"}
    }
  }
}

What it does

MCP Code Execution Server tackles bloated context for agents with dozens of MCP servers. It shows the model one tool, run_python, and three discovery helpers: `discovered_servers()`, `query_tool_docs()` and `search_tool_docs()`. The system prompt stays around 200 tokens regardless of how many servers are connected, and the model loads the schemas it needs on demand. Code runs in a rootless Podman or Docker container. The network is off, the root file system is read-only and all Linux capabilities are dropped. Memory is capped at 512 MB and processes at 128. Writing is allowed in /tmp and /workspace. The container starts once and stays active, so variables, imports and functions from one call are available in the next. Inside the sandbox there are proxies to your MCP servers: `mcp_filesystem.read_file(...)`, `mcp_serena.search(...)` and others. The server finds them in the configurations of Claude Code, Claude Desktop, Cursor, OpenCode, Windsurf and VS Code, and you pass the list of servers to load in the servers parameter.

Who it is for. For developers and analysts who connected many MCP servers to Claude Code or Cursor and want to save context while getting isolated Python execution.

Example request

Through run_python find all TODOs in the project with the filesystem tool, group them by file with pandas and print a table

MCP

Transport

stdio

Authentication

not required

Limitations

You need Python 3.11 or newer, Podman or Docker and the python:3.13-slim image, which has to be pulled in advance. The server starts through uvx straight from the GitHub repository. The default call timeout is 30 seconds with a 120-second maximum, and the container shuts down after five minutes of idleness.

Target servers are discovered in client configurations automatically, and the bridge's own entry in those configurations is skipped to avoid a recursive launch. The license is GPL-3.0.

How to disable. Remove the mcp-server-code-execution-mode entry from the MCP client configuration and restart the client. The container exits after idling, and you can clear the `~/MCPs` state folder by hand.

README in short

The core idea is «code instead of tool calls», and the author cites material on it from Anthropic, Cloudflare, Docker and Apple. Container constraints, environment variables and server discovery locations are laid out in tables.