Skip to content
Install

Greenhouse

Greenhouse runs coding agents on cloud machines that your Orchard instance manages, so agents work away from your laptop and keep working after you close it. It is a first-party graft, an extension that an operator installs on the instance. Greenhouse gives you an isolated Linux machine with your repositories on it. Each agent session works on its own branch in its own git worktree, which means several sessions can run on one machine at the same time without touching each other’s files. You review the changes in a browser workspace, or in a desktop editor over SSH.

Greenhouse is for writing code. To serve traffic, deploy the code to Orchard as an app.

Greenhouse is available when an operator installs the greenhouse graft on your instance. If the instance has PostHog configured, your account also needs the greenhouse feature flag.

  • Get started: Create a machine and start your first session.
Term Definition
Machine An isolated Linux environment with a persistent home directory. It holds your repositories and sessions. The person who creates it is its host.
Session One piece of work: a git worktree on a new branch, an agent working in it, and its diff, pull request, and preview.
Chat One conversation with the agent in a session. A session can have several chats. All of them share the session’s worktree.
Runner The system a machine runs on: Kubernetes or Vercel. You choose the runner when you create the machine.

The API, the CLI, and the Model Context Protocol (MCP) tools call a machine a sandbox. For example, orchard sandbox list lists your machines.

A session runs one of these agents. You select the agent in the model menu of the chat.

Agent CLI value Credentials
opencode opencode, acp:opencode Your Claude token and OpenAI key, if set, and opencode’s free models
Claude Code acp:claude Your Claude token
Codex acp:codex A ChatGPT sign-in or an OpenAI key
fx acp:fx A Vercel AI Gateway, ChatGPT, or Grok sign-in, or an AI Gateway key

The default agent is opencode. An operator can change the default with SANDBOX_DEFAULT_AGENT.

An instance offers one or both runners.

Runner A machine is A stopped machine keeps Previews
Kubernetes A pod on the cluster, in a Kata VM if the operator configured Kata Its home volume Private, through the Orchard server
Vercel A Vercel Sandbox on the operator’s Vercel account Its drive and a filesystem snapshot Public URLs that anyone with the link can open

Machines describes what each runner keeps and which limits apply.

A machine holds the credentials of its host, including agent tokens and GitHub access. Share a machine only with people you trust with those credentials.

On the Kubernetes runner, a machine has internet egress only. It cannot reach cluster services, other machines, or cloud metadata. Nothing outside the cluster can connect to it directly. SSH, terminals, and previews go through the Orchard server.