A Claude Code plugin that restricts Claude to ONLY use Python code execution. All other tools are blocked, creating a pure computational environment.

---

What This Does

• ✅ Allows: execute_code - One Tool to Rule Them All. Execute Python code and capture results.
• ❌ Blocks: All other tools (Read, Write, Web Fetch, Edit, Bash, Grep, Glob, etc.)
• 📚 Includes: Built-in skill that educates Claude about the restricted environment

Claude can only run Python code and is forced to write and execute code to do any real work. This creates a constrained environment ideal for deterministic, computational tasks.

Blog Post

Installation

Note: This is experimental. A demo. It may not work for you. It is not production-ready. Use at your own risk.

git clone https://github.com/rvantonder/execute_code_py.git

cd execute_code_py
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt

claude --plugin-dir .

Now your Claude session can ONLY execute Python code!

Examples

You: randomly generate 1000000 numbers between 0 and 100000 and sort them

Claude

⏺ code-executor - execute_code (MCP)(code: "...")
  ⎿  PreToolUse:mcp__code-executor__execute_code hook succeeded: Success
  ⎿                                                                                                                                                              
     import numpy as np
     import time

     # Generate 1,000,000 random numbers between 0 and 100,000
     gen_start = time.time()
     numbers = np.random.randint(0, 100001, size=1000000)
     gen_time = time.time() - gen_start

     # Sort the numbers
     sort_start = time.time()
     sorted_numbers = np.sort(numbers)
     sort_time = time.time() - sort_start
   ...

You: Create a bar chat of a binomial distribution in svg

...

Internals and Known Limitations

Claude Code is not inherently set up to enable what Code-Only aspires to. For example:

• calling a custom tool needs to go through an unnecessary MCP abstraction
• we have to use PreHook tooling to restrict Claude from calling other tools.

Native support would provide a much better UX:

• elide parameters that the code is called with (currently this is hardcoded to display in Claude Code) and better highlighting.
• code execution in the same runtime as the agent (🤩)

Nevertheless, the concept stands!

How it works--

1. MCP Server

Exposes the execute_code tool with:

• Input: Python code string + optional working directory
• Output: Execution results, stdout, stderr, and error messages
• Features: Captures result variable, handles large outputs to temp files

2. Tool Enforcement Hook

A PreToolUse hook that:

• Intercepts ALL tool calls before execution
• Allows ONLY mcp__code-executor__execute_code
• Denies all other tools with explanation message

3. Environment Education Skill

Teaches Claude about:

• The restricted environment and available capabilities
• How to work effectively with only code execution
• Best practices for computational problem-solving

Tool Definitions

See these as tweakable harness parameters.

Schema

{
  "name": "execute_code",
  "description": "Execute Python code. Set a 'result' variable to return data.",
  "parameters": {
    "code": {
      "type": "string",
      "description": "Python code to execute",
      "required": true
    },
    "working_dir": {
      "type": "string",
      "description": "Working directory (default: current directory)",
      "required": false
    }
  }
}

Response Format

{
  "success": true/false,
  "result": "<value of result variable>",
  "stdout": "<captured output>",
  "stderr": "<captured errors>",
  "result_file": "<path if result too large>",
  "error": "<error message if failed>"
}