agent-jail-mcp

Agent Jail gives AI agents disposable, isolated Docker sandboxes.

Instead of letting agents run commands directly on your machine, Agent Jail creates ephemeral containers where agents can safely execute code.

Give your AI agent a jail, not your machine.

Features

• MCP server for AI clients

• REST API

• Web dashboard

• Disposable Docker sandboxes

• TTL-based automatic cleanup

• Memory limits

• Optional network isolation

• Command execution inside sandboxes

• Simple sandbox lifecycle management

Related MCP server: Containerized Strands Agents

Components

MCP Server

Exposes sandbox tools to MCP-compatible clients.

python server.py

REST API

python api.py

Smoke Test

python smoke_test.py

Docker

Build and run the complete stack:

docker compose up --build

Requirements

• Python 3.12+

• Docker

• Docker Compose

Setup

Create a virtual environment:

python3 -m venv venv
source venv/bin/activate

Install dependencies:

pip install -r requirements.txt

Copy the environment configuration:

cp .env.example .env

Configure .env as needed.

Architecture

The MCP server receives sandbox requests from an AI agent and delegates sandbox lifecycle and command execution to Docker.

Each sandbox is disposable and isolated from the host environment.

Security

Agent Jail is designed to reduce the risk of allowing AI agents to execute commands directly on the host.

Sandboxes should be treated as untrusted execution environments.

For production deployments, carefully review Docker permissions, resource limits, network access, mounted volumes, and container capabilities.

License

MIT