Paddock

Product Introduction

1. Definition: Paddock is a free, open-source, MIT-licensed self-hosted virtualization and containerization platform designed specifically for AI agents. Categorized under developer tooling, AI infrastructure, and secure sandboxing solutions, it runs on a single Virtual Private Server (VPS) owned by the user. The platform provides each AI agent with either a dedicated Linux computer (complete with an Ubuntu XFCE desktop environment and persistent storage) or a locked-down, ephemeral code execution sandbox.

2. Core Value Proposition: Paddock exists to solve the infrastructure and security gaps of deploying autonomous AI agents that require desktop GUIs, persistent file systems, or strict network isolation. By combining Docker, Sysbox container virtualization, and a comprehensive CLI, API, and TypeScript/Python SDK suite, Paddock enables developers to host, monitor, and automate AI agent workflows on affordable, self-owned hardware while retaining total data sovereignty and cost predictability.

Main Features

1. Computer Use and Desktop Environments: Paddock provisions full Ubuntu XFCE desktops per agent, allowing multimodal AI agents (such as Claude Code, Codex, OpenCode, Pi, Prime, and Kimi) to open GUI applications, interact with software visually, and capture screenshots via the CLI (paddock screenshot) to verify build outputs. Preinstalled agents can execute tasks natively using integrated skills and terminal commands.

2. Secure Sandboxing and Network Controls: The platform delivers fine-grained security policies through network isolation modes, including allowlist filtering, complete denial (deny-all), and unrestricted access. Cloud metadata endpoints are blocked by default to prevent credential exfiltration. Additionally, Sysbox integration allows Docker containers to run securely nested inside the box without privileged container vulnerabilities.

3. Snapshots, Drives, and Persistence: Paddock automatically captures box snapshots every minute and upon stop events, enabling instant rollback, pausing, resuming, and branching via paddock restore. Persistent data management is handled through "Drives"—storage volumes under /mnt that mount read-write into a single active box or read-only across multiple concurrent boxes, ensuring files outlive ephemeral agent sessions.

Problems Solved

1. Pain Point: Running autonomous AI agents locally or on managed cloud infrastructure often exposes host systems to arbitrary code execution risks, lacks native GUI interaction for desktop-based tasks, and incurs unpredictable cloud scaling costs. Furthermore, stateless agent execution forces repetitive environment setup, losing generated tools, caches, and intermediate files between runs.

2. Target Audience: AI Engineers, Software Developers building agentic workflows, DevOps Engineers managing LLM infrastructure, and QA Automation Specialists deploying autonomous testing agents.

3. Use Cases: Automated bug fixing and test execution in isolated CI/CD pipelines; web scraping and documentation generation using headless or GUI browsers; executing untrusted third-party code in locked-down sandboxes; and long-running software development tasks where AI agents require persistent home directories, custom toolchains, and visual verification of user interfaces.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros: Full self-hosting control on any Ubuntu/Debian VPS; zero licensing costs due to the MIT open-source license; built-in desktop GUI support for visual agent verification; automated minute-by-minute snapshotting; granular network egress controls; and preinstalled support for leading coding agents like Claude Code and OpenCode.
   • Cons: Requires manual initial setup on an Ubuntu/Debian VPS with ports 80 and 443 open; limited out-of-the-box cluster orchestration since it is explicitly engineered for a single VPS; lacks native support for non-Linux guest operating systems.

2. Key Alternatives & Differentiation:

   • Docker Containers (Raw): While raw Docker provides containerization, it lacks built-in XFCE desktop streaming, automated minute-by-minute snapshotting, native multi-user isolation, and pre-configured AI agent runtimes.
   • Modal / Replicate / E2B: These serverless cloud platforms offer secure execution sandboxes for AI code, but they operate on pay-per-execution cloud pricing models, lack dedicated persistent desktop environments with GUI support, and require data to leave private infrastructure. Paddock differentiates itself by running entirely on a single, fixed-cost VPS owned by the user with complete data privacy.

Frequently Asked Questions (FAQ)

1. What operating systems are supported for hosting Paddock? Paddock installs natively on Ubuntu or Debian VPS environments that have ports 80 and 443 open. The client CLI and agent skills can be installed on local developer laptops running macOS, Linux, or Windows.

2. How does Paddock prevent AI agents from accessing sensitive cloud metadata? Paddock automatically blocks cloud metadata service IP addresses and endpoints at the network level for every box, preventing agents from exploiting cloud IAM roles or instance credentials.

3. Can multiple AI agents share the same storage volume or workspace? Yes. Paddock utilizes "Drives," which are persistent folders mounted under /mnt. A drive can be mounted as read-write into a single active box or shared as read-only across multiple concurrent boxes to facilitate collaboration and shared asset access.

4. Which AI coding agents come preinstalled in Paddock boxes? Paddock comes preconfigured with popular AI development agents including Claude Code, Codex, OpenCode, Pi, Prime, and Kimi, which can be invoked directly via the paddock prompt command or CLI integration.