Archives

Keep Pi and your model credentials local while its coding tools run in an isolated Celesto cloud workspace, then safely sync the changes back.

Firecracker does not run natively on macOS. Learn how to develop with SmolVM and QEMU on a Mac, then switch to Firecracker on Linux with one backend flag.

Measured Firecracker and QEMU performance, an honest libkrun data gap, reproducible benchmark commands, and practical sizing profiles for AI sandboxes.

CelestoFS gives AI agent sandboxes a durable petabyte-scale workspace filesystem for repositories, build artifacts, datasets, and workspace state.

Learn how qcow2 backing files work in QEMU, why they are different from Linux overlayfs, and how they make VM sandbox cloning fast and cheap.

SmolVM now supports Windows 11 sandboxes. Boot disposable Windows VMs from Python, run PowerShell, upload files, pass environment variables, and tear the VM down after the task.

SmolVM is an open-source microVM runtime that lets AI agents run untrusted code in a hardware-isolated sandbox. Learn why Docker isn't enough for LLM-generated code and how to spin up a Firecracker-powered sandbox in three lines of Python.

Not all isolation is equal. Here's how to pick the right one for your AI agents.

Give it a disposable computer, not your laptop.

Give it a disposable computer, not your laptop.

Build an agentic personal chef that learns your preferences over time.

Build an AI agent that scrapes Pokemon card prices from Cardmarket using Agentor and ScrapeGraphAI. Full code included.

I cancelled ChatGPT after switching to OpenClaw—an always-on assistant you can use via Telegram, WhatsApp, or iMessage. Here’s the difference.

Running OpenClaw on your laptop is convenient—but unreliable and risky. Learn the real failure modes and safer ways to host it (VPS, sandboxing, or Celesto).