mcp-python-exec-sandbox

Sandboxed Python execution for AI agents. Scripts run in ephemeral, isolated environments with inline dependencies (PEP 723) -- zero host pollution, zero leftover venvs, zero package conflicts.

Why?

Every coding agent can already run Python on your host. The problem is what happens next: packages accumulate, venvs sprawl, and a rogue pip install breaks your system. mcp-python-exec-sandbox eliminates this:

• Scripts execute in a sandbox (bubblewrap on Linux, Docker on macOS/other platforms)
• Dependencies are declared inline and resolved ephemerally via uv
• Nothing touches your host's Python, site-packages, or virtualenvs
• Each execution is isolated and disposable

Features

• Sandboxed execution -- platform-specific isolation prevents host filesystem access
• PEP 723 inline metadata -- declare dependencies directly in scripts with # /// script blocks
• Multi-version Python -- run scripts on Python 3.13, 3.14, or 3.15 (uv downloads the right version automatically)
• Ephemeral environments -- dependencies are resolved per-execution, never persisted
• Package caching -- uv's global cache makes repeat installs near-instant
• Timeout enforcement -- configurable per-execution timeouts
• Output truncation -- prevents runaway output from overwhelming the agent

Prerequisites

All setups require:

• Python 3.13+ -- to run the MCP server process
• uv -- manages script execution, dependency resolution, and Python version downloads. Also provides uvx for running the server without installing it globally.

Additional requirements depend on your chosen sandbox backend:

Host Python vs. execution Python: These are independent. Python 3.13+ is needed to run the server process itself. The --python-version flag controls which Python version your scripts execute on -- uv downloads the target version automatically. You do not need to install Python 3.14 or 3.15 on your host to run scripts on those versions.

Quick start

Claude Code (Linux -- native sandbox)

claude mcp add python-sandbox -- uvx mcp-python-exec-sandbox

Claude Code (macOS -- Docker sandbox, recommended)

claude mcp add python-sandbox -- uvx mcp-python-exec-sandbox

The Docker sandbox image is pulled automatically from GHCR on first use. No manual build required.

Claude Code (no sandbox)

claude mcp add python-sandbox -- uvx mcp-python-exec-sandbox --sandbox-backend none

Cursor

Add to .cursor/mcp.json (project-level) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "python-sandbox": {
      "command": "uvx",
      "args": ["mcp-python-exec-sandbox"]
    }
  }
}

OpenAI Codex CLI

codex mcp add python-sandbox -- uvx mcp-python-exec-sandbox

Or add to .codex/config.toml:

[mcp_servers.python-sandbox]
command = "uvx"
args = ["mcp-python-exec-sandbox"]

Other MCP clients

Any client that supports the MCP stdio transport can use this server:

{
  "mcpServers": {
    "python-sandbox": {
      "command": "uvx",
      "args": ["mcp-python-exec-sandbox"]
    }
  }
}

Multi-version Python

Use --python-version to target a specific Python version. uv downloads it automatically -- no manual install needed.

# Python 3.13 (default)
uvx mcp-python-exec-sandbox --python-version 3.13

# Python 3.14
uvx mcp-python-exec-sandbox --python-version 3.14

# Python 3.15
uvx mcp-python-exec-sandbox --python-version 3.15

This works across all sandbox backends. The Docker sandbox uses uv inside the container to manage Python versions, so the same --python-version flag applies.

Tools

execute_python

Execute a Python script with automatic dependency management.

# Simple script
execute_python(script="print('hello world')")

# Script with dependencies
execute_python(
    script="import requests; print(requests.get('https://httpbin.org/get').status_code)",
    dependencies=["requests"]
)

# Script with inline PEP 723 metadata
execute_python(script="""
# /// script
# dependencies = ["pandas", "matplotlib"]
# ///

import pandas as pd
print(pd.DataFrame({'a': [1,2,3]}).describe())
""")

check_environment

Returns information about the execution environment: Python version, uv version, platform, sandbox status, and configuration.

validate_script

Validates a script's PEP 723 metadata and dependencies without executing it.

Sandbox backends

The default backend is native (bubblewrap) on Linux and docker on macOS/other platforms. Specifying --sandbox-backend native on macOS automatically redirects to Docker. If the sandbox tool is unavailable, the server falls back to none with a warning.

Docker sandbox setup

The Docker sandbox image is published to GHCR and pulled automatically when the server starts. No manual setup is needed.

To build locally for development:

docker build -t ghcr.io/lu-zhengda/mcp-python-exec-sandbox profiles/

CLI options

mcp-python-exec-sandbox [OPTIONS]

Options:
  --python-version TEXT     Python version for execution (default: 3.13)
  --sandbox-backend TEXT    native | docker | none (default: native on Linux, docker on macOS)
  --max-timeout INT         Maximum allowed timeout in seconds (default: 300)
  --default-timeout INT     Default timeout in seconds (default: 30)
  --max-output-bytes INT    Maximum output size in bytes (default: 102400)
  --no-warm-cache           Skip cache warming on startup
  --uv-path TEXT            Path to uv binary (default: uv)

Development

Setup

git clone https://github.com/lu-zhengda/mcp-python-exec-sandbox.git
cd mcp-python-exec-sandbox
uv sync --dev

Project structure

src/mcp_python_exec_sandbox/   # Package source
  server.py               # FastMCP server + tool definitions
  executor.py             # uv subprocess orchestration
  script.py               # PEP 723 metadata parsing/merging
  sandbox.py              # Sandbox ABC + factory
  sandbox_linux.py        # bubblewrap sandbox (Linux)
  sandbox_docker.py       # Docker sandbox (macOS/any)
  config.py, cache.py, output.py, errors.py
tests/                    # Unit + integration tests (mocked or local uv)
e2e_tests/                # End-to-end tests (require uv + network)
profiles/                 # Dockerfile, warmup packages
.devcontainer/            # Devcontainer for Linux sandbox testing from macOS

Running tests

Unit and integration tests -- fast, run everywhere:

uv run pytest tests/ -v

E2E tests -- require uv and network access. These exercise real script execution, package installation, MCP protocol flow, and sandbox enforcement:

uv run pytest e2e_tests/ -v

Docker sandbox tests

The Docker E2E tests (e2e_tests/test_docker_sandbox.py) verify execution, dependency installation, read-only filesystem enforcement, host isolation, and timeout handling through the Docker backend.

Prerequisites:

1. Docker must be installed and running
2. Build the sandbox image:

docker build -t ghcr.io/lu-zhengda/mcp-python-exec-sandbox profiles/

Then run:

uv run pytest e2e_tests/test_docker_sandbox.py -v

These tests are automatically skipped if Docker is unavailable or the image hasn't been built.

Linux sandbox tests (devcontainer)

The Linux sandbox tests (e2e_tests/test_sandbox_enforcement.py::test_linux_sandbox_blocks_etc_shadow) use bubblewrap (bwrap) for namespace isolation. They are skipped on macOS because bwrap is Linux-only.

To run them from macOS, use the included devcontainer which provides Ubuntu 24.04 with bwrap pre-installed:

VS Code:

1. Install the Dev Containers extension
2. Open the project and select Reopen in Container
3. In the integrated terminal:

uv run pytest e2e_tests/test_sandbox_enforcement.py -v

CLI: