← In the News

Overalls puts coding agents on self-hosted microVMs and keeps API keys out of them

Overalls: an AI-first developer platform that you own · Project site, read October 16, 2026

Machine-readable Download Markdown

The page describes a self-hosted system where each development machine is a persistent Firecracker microVM running NixOS, Ubuntu or Windows Server 2025. A machine keeps its disk, name and address when stopped, resized, cloned or moved. It ships with a built-in agent that works in threads on a machine, wakes it if it is stopped and keeps running after the laptop closes. The page calls this "the Amp orbs workflow on hardware you own." Claude Code and Codex come pre-installed. The site says it is MIT-licensed and that there is no hosted version.

The design detail that matters for agent work is how credentials are handled. Guests reach OpenAI and Anthropic models through an internal address, and a proxy adds the provider credentials outside the VM, so no API key lives on the machine. Named integrations work the same way: the page says a machine can use a secret without being able to read it. Usage is recorded per user, down to the machine and agent thread, and the page says unmeasured usage shows as "unknown, never as zero." The page also describes a locked-down profile with no outbound network, no sharing and SSH for the owner only, intended for customer material.

Two limits on what we can say. The stability labels (most features are marked "STABLE", with several guest operating systems marked "EXPERIMENTAL") are the project's own, and we found no independent testing. The page links to a GitHub repository under the ghuntley account and to ghuntley.com, but does not name an author, and the repository link returned a GitHub not-found page when we opened it, so the open-source claim and the code behind it are unconfirmed.

Why it matters: Putting the credential outside the machine and metering by agent thread are two things a team running agents on shared infrastructure has to solve either way, and this is a concrete example of one way to do it. Whether it holds up depends on code we could not open, so treat it as a design to compare against, not a platform to adopt yet.