Caterm

REMOTE AI AGENTS

Codex & Claude Code.
Your machine. Your browser.

Caterm lets you send prompts to Codex or Claude Code from a browser while the agent CLI executes on a connected Mac or Linux computer. It uses the device's existing local login, project files, and tools.

What is a remote AI agent in Caterm?

A remote AI agent is a locally installed coding agent you interact with through Caterm. Choose a device, agent, and project when creating a chat. Caterm relays messages, tool activity, questions, approval requests, and stop commands between your browser and that device.

This is separate from remote desktop control: agent conversations use a chat interface and do not require a desktop stream or a Linux graphical desktop. You can use the browser terminal or file browser alongside the agent when those capabilities are enabled.

Set up your agent

  1. Connect a computer with the Caterm Mac app or Linux daemon.
  2. Install and sign in to Codex or Claude Code on that computer. Caterm does not include an agent provider subscription or model usage.
  3. Configure the agent executable and project root locally, then open Chat → New conversation in Caterm.
  4. Choose a name, online device, agent, and relative project folder. Send a prompt and follow the response and tool activity.

Read the Codex and Claude Code configuration guide →

Permissions and human input

Codex chats offer Approve for me using Codex's native risk reviewer, or Full access. Ordinary Claude Code chats use native manual permissions. Respond to forwarded approval requests or questions in the conversation, and stop a turn when needed.

The project folder is not an agent sandbox

Project selection restricts where the session starts. The agent's native execution policy controls tool access. Full access can permit commands, network access, and files beyond that project. Caterm's scoped file browser has a separate boundary.

What happens if I close the browser?

A running agent turn continues on the device when the browser disconnects. Reopen the conversation to receive persisted events without rerunning the prompt. The computer must stay awake, online, and connected for remote interaction; a daemon restart ends interrupted turns instead of automatically relaunching them. Runs have timeouts and can pause for your input.

Follow-up messages resume the native session for that Caterm conversation. Caterm does not import an arbitrary session already open in your local terminal. A conversation's device, agent, and project remain fixed; create a new chat to change them.

Where does the data go?

Commands and agent tools execute on the device. Caterm relays messages and stores conversation history on its server. The agent communicates with its model provider according to its local configuration. Local execution does not mean offline AI, a locally hosted model, or that all data stays on the computer.

Sharing and automation

Device sharing lets other verified accounts use enabled tools and create separate chats. The device's local agent login, files, operating-system user, resources, and provider quota remain shared.

For repeatable work, create a scheduled workflow with ordered prompts. Every run creates its own native session on the selected device.