Overview

A local-first workbench that runs coding agents in sealed, throwaway microVMs.

Isolade runs coding agents (Claude Code, Codex, or a plain shell) inside isolated VM sandboxes on your own machine. You create workspaces, spin up instances, and talk to the agents through a desktop app.

An agent gets far more useful once you stop approving every command it runs. The catch is that a free-roaming agent on your laptop can read your SSH keys, your tokens, and every repo in your home directory. Isolade removes that risk by giving each agent its own microVM — a hardware boundary with none of your files or real secrets inside it — so you can hand it the whole task and walk away.

What you get

Each agent runs in its own microVM, built on Apple’s Hypervisor framework or KVM. That’s a hardware boundary, so an escape has to beat the CPU instead of a syscall filter.

Secrets never enter the VM. The agent sees a placeholder, and the real value is spliced into outgoing requests at the proxy, only toward hosts you’ve named. The agent can push to GitHub without ever holding your token.

Because the VM is sealed and disposable, the agent runs without approval prompts. It reads, writes, and executes whatever the task needs, and the worst case is a VM you delete.

Your source tree stays on the host and is bundled into local VMs. Model providers receive the prompts and task context the agent sends during normal API calls.

A workspace is config in git: one short TOML file and a Dockerfile. Commit it, and a teammate gets the same environment, down to the build cache.

Where to go next

• Quickstart: install, build a workspace, run your first agent.
• Workspaces: the config.toml reference.
• Secrets: how the placeholder substitution works.
• How it works: the path from a config file to a running agent.
• Architecture: packages, the builder VM, and the HTTP API.

Requirements