MCP Server Code Execution Mode

Skill by ara.so — MCP Skills collection.

This MCP server implements Anthropic's "Code Execution with MCP" pattern, exposing a single run_python tool instead of hundreds of individual tools. The agent writes Python code to discover, call, and compose MCP tools dynamically, reducing context overhead from ~30,000 tokens to ~200 tokens.

Key capabilities:

• Execute Python in rootless Podman/Docker containers
• Proxy any stdio MCP server into the sandbox
• Discover tools at runtime (no context preloading)
• Fuzzy search across all connected servers
• Persistent sessions (variables/state retained)
• Security: no network, read-only filesystem, dropped capabilities

Installation

Prerequisites

Install a container runtime (rootless mode):

# macOS (Podman Desktop recommended)
brew install podman
podman machine init
podman machine start

# Linux
sudo apt install podman  # or dnf/yum/pacman
podman system migrate    # enable rootless

# Windows (WSL2 + Podman Desktop)
# Download from https://podman-desktop.io/

Install the Bridge

# Via pip
pip install mcp-code-execution

# Via uv (recommended)
uv pip install mcp-code-execution

# From source
git clone https://github.com/elusznik/mcp-server-code-execution-mode.git
cd mcp-server-code-execution-mode
uv pip install -e .

Pull the Container Image

# Pre-built image (recommended)
podman pull ghcr.io/elusznik/mcp-code-execution:latest

# Or build custom image
podman build -t mcp-code-execution:custom -f Dockerfile .

Configuration

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "code-execution": {
      "command": "uv",
      "args": [
        "--directory",
        "/path/to/mcp-server-code-execution-mode",
        "run",
        "mcp-code-execution"
      ],
      "env": {
        "MCP_BRIDGE_RUNTIME": "podman",
        "MCP_BRIDGE_IMAGE": "ghcr.io/elusznik/mcp-code-execution:latest",
        "MCP_BRIDGE_OUTPUT_MODE": "compact"
      }
    }
  }
}

Environment Variables

Proxying Other MCP Servers

Create a mcp_bridge_config.json in your working directory:

{
  "servers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
      "env": {}
    },
    "github": {
      "command": "uvx",
      "args": ["mcp-server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
      }
    },
    "postgres": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "POSTGRES_CONNECTION_STRING",
        "mcp/postgres"
      ],
      "env": {
        "POSTGRES_CONNECTION_STRING": "${DATABASE_URL}"
      }
    }
  }
}

The bridge auto-discovers configs from:

• ./mcp_bridge_config.json
• ~/.config/mcp-bridge/config.json
• MCP_BRIDGE_CONFIG env var (JSON string)
• Claude Desktop config (proxies sibling servers)

Usage Patterns

Basic Code Execution

# Simple calculation
result = 42 * 1.5
print(f"Answer: {result}")

# Data analysis
import pandas as pd
import numpy as np

data = pd.DataFrame({
    'x': np.random.randn(100),
    'y': np.random.randn(100)
})

correlation = data['x'].corr(data['y'])
print(f"Correlation: {correlation:.3f}")

Discovering MCP Servers

from mcp import runtime

# List all available servers
servers = await runtime.discovered_servers()
for server in servers:
    print(f"- {server}")

Querying Tool Schemas

from mcp import runtime

# Get all tools from a specific server
github_tools = await runtime.query_tool_docs("github")

for tool_name, schema in github_tools.items():
    print(f"{tool_name}: {schema.get('description', 'No description')}")

Fuzzy Tool Search

from mcp import runtime

# Search across all servers
matches = await runtime.search_tool_docs("create issue", limit=5)

for hit in matches:
    print(f"{hit['server']}.{hit['tool']}: {hit.get('description', '')}")
    print(f"  Score: {hit.get('score', 0):.2f}")

Calling MCP Tools Directly

# Dynamic lookup
result = await mcp_servers["github"].call_tool(
    "create_issue",
    {
        "owner": "elusznik",
        "repo": "mcp-server-code-execution-mode",
        "title": "Add feature X",
        "body": "Description here"
    }
)
print(result)

# Attribute access
result = await mcp_github.create_issue(
    owner="elusznik",
    repo="mcp-server-code-execution-mode",
    title="Bug report",
    body="Steps to reproduce..."
)

# Module import pattern
from mcp.servers.github import create_issue

issue = await create_issue(
    owner="myorg",
    repo="myrepo",
    title="Task",
    body="Details"
)

Composing Multiple Tools

from mcp import runtime

# 1. Search for calendar tools
cal_tools = await runtime.search_tool_docs("calendar events", limit=3)
calendar_server = cal_tools[0]["server"] if cal_tools else None

if calendar_server:
    # 2. Get today's events
    events = await mcp_servers[calendar_server].call_tool(
        "list_events",
        {"date": "2025-01-15"}
    )
    
    # 3. Create GitHub issues for each event
    for event in events:
        await mcp_github.create_issue(
            owner="myorg",
            repo="tasks",
            title=f"Follow-up: {event['title']}",
            body=f"From calendar: {event['description']}"
        )
    
    print(f"Created {len(events)} issues")

Error Handling

from mcp import runtime

try:
    result = await mcp_servers["github"].call_tool(
        "get_issue",
        {"owner": "invalid", "repo": "repo", "issue_number": 999}
    )
except Exception as e:
    print(f"Tool call failed: {e}")
    
    # Fallback: search for alternative tools
    alternatives = await runtime.search_tool_docs("get issue")
    print(f"Found {len(alternatives)} alternative tools")

Persistent State

Variables and imports persist across calls in the same session:

# First call
import pandas as pd
df = pd.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 6]})

# Second call (same session)
# df and pandas are still available
print(df.describe())

Advanced Patterns

Building a Tool Catalog

from mcp import runtime
import json

catalog = {}

# Discover all servers
servers = await runtime.discovered_servers()

for server_name in servers:
    # Get all tools for this server
    tools = await runtime.query_tool_docs(server_name)
    catalog[server_name] = {
        tool_name: {
            "description": schema.get("description", ""),
            "parameters": schema.get("inputSchema", {}).get("properties", {})
        }
        for tool_name, schema in tools.items()
    }

# Save catalog
with open("/tmp/tool_catalog.json", "w") as f:
    json.dump(catalog, f, indent=2)

print(f"Cataloged {len(catalog)} servers")

Conditional Tool Execution

from mcp import runtime

# Search for weather tools
weather_tools = await runtime.search_tool_docs("weather forecast", limit=1)

if weather_tools:
    server = weather_tools[0]["server"]
    tool = weather_tools[0]["tool"]
    
    forecast = await mcp_servers[server].call_tool(
        tool,
        {"location": "San Francisco"}
    )
    
    # If rain predicted, create calendar reminder
    if "rain" in forecast.lower():
        cal_tools = await runtime.search_tool_docs("create event")
        if cal_tools:
            await mcp_servers[cal_tools[0]["server"]].call_tool(
                cal_tools[0]["tool"],
                {
                    "title": "Bring umbrella",
                    "date": "2025-01-16",
                    "time": "08:00"
                }
            )

Batch Processing

import asyncio

repos = ["repo1", "repo2", "repo3"]
results = []

for repo in repos:
    try:
        issues = await mcp_github.list_issues(
            owner="myorg",
            repo=repo,
            state="open"
        )
        results.append({"repo": repo, "count": len(issues)})
    except Exception as e:
        results.append({"repo": repo, "error": str(e)})

for r in results:
    if "error" in r:
        print(f"{r['repo']}: ERROR - {r['error']}")
    else:
        print(f"{r['repo']}: {r['count']} open issues")

Data Science Workflow

import pandas as pd
import matplotlib.pyplot as plt
from io import StringIO

# Fetch data from MCP tool
csv_data = await mcp_filesystem.read_file(path="/tmp/sales.csv")

# Process
df = pd.read_csv(StringIO(csv_data))
summary = df.groupby("region")["sales"].sum().sort_values(ascending=False)

# Generate chart
fig, ax = plt.subplots()
summary.plot(kind="bar", ax=ax)
plt.title("Sales by Region")
plt.tight_layout()
plt.savefig("/tmp/sales_chart.png")

# Save results
report = f"""
Sales Analysis
==============
Total Sales: ${df['sales'].sum():,.2f}
Top Region: {summary.index[0]} (${summary.iloc[0]:,.2f})

Chart saved to /tmp/sales_chart.png
"""

await mcp_filesystem.write_file(
    path="/tmp/sales_report.txt",
    content=report
)

print(report)

Troubleshooting

Container Runtime Not Found

Error: RuntimeNotFoundError: Neither podman nor docker found

Solution:

# Install Podman
brew install podman  # macOS
sudo apt install podman  # Linux

# Or install Docker
brew install docker  # macOS (with Docker Desktop)

Permission Denied (Rootless)

Error: Error: creating container: mkdir /run/user/1000: permission denied

Solution:

# Enable rootless mode
podman system migrate

# Or use Docker rootless
dockerd-rootless-setuptool.sh install

Image Pull Timeout

Error: TimeoutError: Container image pull exceeded 60s

Solution:

# Pre-pull image manually
podman pull ghcr.io/elusznik/mcp-code-execution:latest

# Or increase timeout
export MCP_BRIDGE_STARTUP_TIMEOUT=180

Server Discovery Fails

Error: No servers found via discovered_servers()

Solution:

1. Check config file exists:

   cat ~/.config/mcp-bridge/config.json
   

2. Validate JSON syntax:

   import json
   with open("mcp_bridge_config.json") as f:
       json.load(f)  # Should not raise
   

3. Test server manually:

   npx -y @modelcontextprotocol/server-filesystem /tmp
   # Should output JSON-RPC messages
   

Tool Call Hangs

Error: Tool execution never completes

Solution:

1. Check timeout settings:

   export MCP_BRIDGE_TIMEOUT=600  # Increase to 10 minutes
   

2. Debug in container:

   podman run -it --rm ghcr.io/elusznik/mcp-code-execution:latest /bin/bash
   # Test commands manually
   

Memory Limit Exceeded

Error: OOMKilled or memory allocation errors

Solution:

export MCP_BRIDGE_MEMORY_LIMIT=2g  # Increase to 2GB

Volume Mount Issues (macOS)

Error: Files not accessible in container

Solution:

# Ensure Podman machine has volume sharing enabled
podman machine stop
podman machine set --rootful=false --volume /Users:/Users
podman machine start

Output Format Issues

Problem: Responses too verbose or unstructured

Solution:

# Use compact mode (default)
export MCP_BRIDGE_OUTPUT_MODE=compact

# Or TOON for deterministic tokens
export MCP_BRIDGE_OUTPUT_MODE=toon

Security Best Practices

1. Never expose the bridge directly to untrusted users — the agent can execute arbitrary Python.

2. Use environment variables for secrets:

   {
     "servers": {
       "github": {
         "env": {
           "GITHUB_PERSONAL_ACCESS_TOKEN": "${GITHUB_TOKEN}"
         }
       }
     }
   }
   

3. Restrict filesystem access for proxied servers:

   {
     "servers": {
       "filesystem": {
         "args": ["-y", "@modelcontextprotocol/server-filesystem", "/safe/path"]
       }
     }
   }
   

4. Monitor resource usage:

   podman stats $(podman ps -q --filter ancestor=ghcr.io/elusznik/mcp-code-execution)
   

5. Audit logs: Check container logs for suspicious activity:

   podman logs <container-id>
   

Performance Tips

1. Warm up servers by calling discovered_servers() early in the session.

2. Cache tool schemas instead of querying repeatedly:

   from mcp import runtime
   
   # Cache at session start
   all_tools = {}
   for server in await runtime.discovered_servers():
       all_tools[server] = await runtime.query_tool_docs(server)
   

3. Use attribute access (slightly faster than dynamic lookup):

   # Preferred
   await mcp_github.create_issue(...)
   
   # Slower
   await mcp_servers["github"].call_tool("create_issue", ...)
   

4. Pre-pull images to avoid startup delays:

   podman pull ghcr.io/elusznik/mcp-code-execution:latest
   

References

• GitHub Repository
• Anthropic Blog: Code execution with MCP
• Cloudflare Blog: Code Mode
• MCP Protocol Documentation
• Apple Research: CodeAct