Paddock

Paddock is an open-source platform designed to provide secure, isolated computing environments for AI agents, allowing them to perform real work on full Linux desktops or in protected sandboxes. Each agent receives a dedicated box with persistent files, tools and desktops, all running on a VPS you control. This setup supports a wide range of agent activity, from opening GUI applications and taking screenshots, to executing scripts and running custom code.

The system enables fine-grained control over agent environments, letting users configure network access (full, deny-all, or allowlist), storage, and user permissions. Snapshots and rollback features help preserve state and make recovery from experiments or malicious runs easy. Integration options include a CLI, browser dashboard, HTTP API, and SDKs for Python and TypeScript, giving both developers and operations teams flexible programmatic access.

Paddock is especially geared for those who want to host and monitor AI agents in secure, reproducible sandboxes on their own infrastructure, avoiding cloud lock-in and reducing security risks. Popular agents come pre-installed, and the MIT license allows unrestricted use and modification.

Paddock is best suited for developers, ML engineers, and IT managers seeking to safely deploy, test, or operate autonomous AI agents on infrastructure they control. It's particularly useful for small tech teams and startups running experiments, building agent-based applications, or needing secure sandboxes for untrusted code.

Use cases include agent development, workflow automation, or any scenario where isolated, persistent computing environments for AI tools or agents are required. Security-conscious organizations and those avoiding public cloud dependencies will appreciate the control over network, storage, and execution policies.