Arrakis: How To Build An AI Sandbox From Scratch
AI Engineer World's Fair 2025 · 40:18
AI agent sandbox infrastructure
Arrakis provides self-hosted sandboxes for AI agents to execute code and interact with desktop applications. Developers and agents can use its REST API, py-arrakis Python SDK, or MCP server to create sandboxes, upload files, and run code. Each sandbox runs Ubuntu in a MicroVM, with a VNC server and preinstalled Chrome for computer use; automatic port forwarding makes the graphical interface accessible.
Solo founder Abhishek Bhardwaj built Arrakis after joining South Park Commons and subsequently joined OpenAI to work on Agent Infrastructure. Arrakis uses cloud-hypervisor to run its MicroVMs. Its snapshot-and-restore capability lets agents checkpoint a sandbox and return to its earlier state, including processes and modified files. This supports backtracking through multi-step workflows and exploring alternative paths with Monte Carlo Tree Search.
Arrakis is distributed as software developers run on their own infrastructure, with Linux hosts and KVM virtualization required. Developers can customize sandbox packages through Dockerfiles and supply their own kernel. The project is licensed under AGPLv3, with commercial licensing also offered. Bhardwaj has reported use by iFood, providing a concrete example of organizational adoption.
AI Engineer World's Fair 2025 · 40:18
Affiliations reflect their AIE appearances, not necessarily current employment.
The talk examines Linux namespaces, virtualization, root-filesystem isolation, networking, and VM construction, including work toward subsecond boot times.
Affiliations reflect each recorded session, not necessarily current employment.