The coding agent

exe has a coding agent of its own: say what you want built, and it builds it inside one of your VMs. If you already work with Claude Code, Codex or another agent, exe teaches that agent to use the VMs instead.

What it is for

For you:

• Build in a VM by asking. Type what you want in a VM's Agent tab. The agent installs packages, writes code and starts services there, and you watch it work.
• Run the whole cloud by talking. The Chat window is a conversation that can see and drive every VM: create one, work inside it, publish its port.
• Tasks that outlast the browser. A reply runs in the daemon. Close the tab and it carries on; come back and rejoin it live.
• Work kept inside the VM. The agent is free to install and run what it needs there, and the VM is the sandbox boundary.
• Your choice of model. Ollama, local or cloud, or a ChatGPT subscription.

For a coding agent of your own, once it has read the Agent Skill Guide:

• A sandbox when it needs one. It creates a VM, or starts one that exists, and may install and run anything inside it.
• A shell in the VM. It runs commands, copies files in and tunnels ports over SSH.
• A way to show its work. It finds the ports listening in a VM and publishes one at an HTTPS address.
• Notes to read and to leave. Each VM carries notes, which the agent reads before working in a VM it does not know and can rewrite.
• A line to you. It posts a note to the Newsfeed, on this node and every joined one, when work is finished or something is wrong.
• A helper. It can pass a task on to the built-in agent.

Choose a model

The built-in agent needs a model. Set one up in Windows → Configuration.

Ollama is the default, on the Ollama tab:

The Chat window can run on a ChatGPT subscription instead of Ollama. On the OpenAI tab:

1. Click Sign in with ChatGPT… and finish the sign-in in the browser. It is the OAuth flow the Codex CLI uses, with no API key.
2. Set chat_provider to openai.
3. Pick a model (gpt-5.4, gpt-5.4-codex, …) and, if you like, a reasoning effort.
4. Press Save.

The browser sign-in redirects to localhost:1455, and the daemon listens on all interfaces there. When the daemon runs on another machine, swap localhost for the daemon's host in that final URL, or paste the URL into the tab's paste field.

Tokens live in ~/.exe/openai.json and refresh themselves. While you are signed in, the tab also shows the subscription's rate-limit usage (the rolling 5-hour and weekly windows, with their reset times) and any credit balance.

A VM's Agent tab follows the same choice, since what it starts is a chat: with chat_provider set to openai, its runs are on the ChatGPT subscription too. To change provider, model or effort without opening Configuration, click the Chat window's status line.

Run the agent on a VM

1. Open the VM's window and click its Agent tab.
2. Type what you want built.
3. Press Run Agent.

Run Agent opens a new chat pinned to the VM, with your prompt as its first message. The Chat window streams the run and takes what you say next.

The agent can install packages, write code and start services. It has passwordless sudo (doas on Alpine) inside the VM, and the VM is the sandbox boundary.

The Chat window

The Chat icon and window appear once a chat backend is usable. Chat is a conversation that can see and drive your whole VM cloud: it can create, start, stop and delete VMs, run commands and edit files inside them, publish a port and push a folder to GitHub.

• Replies run in the daemon, not in the browser. Closing the tab, or losing the network, never interrupts a long task.
• To rejoin a reply, reopen the chat and select the session, which is marked with a green dot while it is streaming.
• Stop actually cancels the run. The Send button reads Stop while a reply runs, and Queue once you have typed something: your message then joins the run in progress.
• It asks first. Before Chat deletes a VM or removes a published address, a dialog asks you to approve it.
• A chat pinned to one VM, which is what the Agent tab starts, works in that VM alone.

From the command line

exe code, on the machine exe runs on, runs the agent in a running VM and prints its work as it goes:

exe code demo "build a guestbook app on port 8000"

exe code, and the POST /v1/vms/{name}/agent call behind it, stay on Ollama whatever chat_provider says. So does code in the SSH lobby.

Hand exe to your own agent

Prefer your own agent? exe serves a /skill.md file that teaches Claude Code, Codex or any other coding agent how to drive the API and the VMs.

1. Choose Help → Agent Skill Guide…. The window shows the guide's address and a prompt that tells an agent to read it.
2. Press Copy Prompt.
3. Paste the prompt into your agent.

When this browser holds an API token, the prompt carries it and the window says so. Share that prompt with care.

The guide also gives the agent its limits. It is not to stop or delete a VM it did not create unless you ask, and it publishes to GitHub or changes the desktop menu only when you ask.

The SSH gate an agent works through is in SSH from your own terminal. To run Claude Code or Codex on this machine, in a window that keeps its sessions, see Claude Code, Codex and Terminal.