Skip to main content
Type opencode from any directory on any boxd VM. The CLI is already on the image, reads the standard AGENTS.md, and runs against the same in-VM boxd CLI the other agents use. Real kernel, full root, persistent disk, public URL.

​
How it works

OpenCode is pulled in at image build time, so every fresh VM and every fork has it. The image ships an AGENTS.md at ~/.config/opencode/AGENTS.md covering the boxd environment: how to create siblings with boxd new, manage proxies, exec across VMs, and clean up. OpenCode picks it up by default from any working directory. The in-VM boxd CLI is pre-authenticated by source IP, so the agent can drive the platform without a token. JSON output is on by default with --json, so the agent can parse what it ran. Auth for the model itself is not yet persisted across resets, so first run on a fresh VM means signing in. Claude Code on boxd already persists auth across resets and forks (how); the same UX is on the roadmap for OpenCode.

​
Use it

Interactive session: 
ssh myapp.boxd.sh
cd my-project
opencode
Non-interactive delegation through boxd exec: 
boxd exec myvm 'opencode run "Add a /health endpoint to the Flask app"'
The same sandbox patterns documented for the other agents apply: fork before risky ops, fan out across VMs, destroy when done. See Agent sandboxes.

​
Patterns

​
Fan-out across VMs

for i in 1 2 3; do boxd new --name=try-$i --json & done; wait

boxd exec try-1 'opencode run "Use FastAPI" &'
boxd exec try-2 'opencode run "Use Flask" &'
boxd exec try-3 'opencode run "Use Starlette" &'

# Results: https://try-1.boxd.sh, try-2, try-3

​
Fork before risky ops

Snapshot a working VM before a destructive change. If the run goes wrong, destroy the fork, fork again from the parent. 
boxd fork myvm --name=myvm-refactor --json
boxd exec myvm-refactor 'opencode run "Refactor the auth module"'

​
Bias with project-level AGENTS.md

Drop an AGENTS.md in your repo. OpenCode merges it with the boxd-provided one, so you keep the platform context for free and add your conventions on top.
Auth doesn’t persist across resets for OpenCode yet. Plan to sign in once per VM, or use Claude Code on boxd for the persistent-auth UX.

​
FAQ

~/.config/opencode/AGENTS.md is the global file the image ships with. A repo-level AGENTS.md is merged on top when you run opencode in that directory.
Yes. The in-VM boxd CLI works from any agent. OpenCode can boxd new siblings, boxd exec into them, and boxd destroy when done.
Not yet. The slash-command skill targets Claude Code today. Email contact@boxd.sh and we’ll prioritize.
Yes. The VM is a real KVM microVM with its own kernel. Anything you’d do on a Linux laptop, OpenCode can do here.

​
Next

https://mintcdn.com/azin/Ax1V0serIwQf0x_2/images/icons/robot.svg?fit=max&auto=format&n=Ax1V0serIwQf0x_2&q=85&s=e91e744c2d3cd5da053167486b6834bf

Agent sandboxes

Give OpenCode its own VM. Fork to recover.
https://mintcdn.com/azin/Ax1V0serIwQf0x_2/images/icons/claude.svg?fit=max&auto=format&n=Ax1V0serIwQf0x_2&q=85&s=03286da259893b6993db73a8e038c20b

Claude Code on boxd

Same primitives, persistent auth, ships with slash-command skills.