Radial app

Agents & terminal

Radial's agent surface is the real thing: the actual Claude Code, Codex, or Cursor CLI running in a real terminal inside your project, with the live preview next to it. No wrapper chat, no separate model bill. The same agent you'd run yourself, with much better aim.

Agent terminals

⌘I opens the agent dock; ⌘T starts a new Claude Code agent in the open project (Codex, Cursor, and a plain shell are in the Agent menu). Each agent is a real PTY spawned in the repo's directory, so the CLI behaves exactly as it does in your own terminal.

• Run as many agents per project as you like, and rename them to keep parallel work straight. They name themselves from their current task by default.
• Sessions are persistent: agents keep running if you close Radial, and reattach with their scrollback when you come back.
• Idle Claude Code sessions are parked when their project sleeps or Radial quits, and come back in the same conversation when you return. To keep one running no matter what, choose Keep awake from its menu in the sidebar rail: Radial never parks it, and your Mac won't idle-sleep while it runs, so a session you drive from your phone with Claude Code's remote control stays reachable. (Closing a MacBook's lid still sleeps it.) Let it sleep in the same menu releases it.
• The dock sits on the right by default; Dock bottom in its header flips it. Terminal, Comments, and History share the dock as tabs.

The terminal is the chat

There is no separate chat box: you type to the agent in its own CLI (⌘L focuses it). Everything Radial hands over lands in the agent's input for you to review and press Enter, so nothing runs without your say:

• Elements: capture with the Select tool and the exact selector, text, and page arrive in the input. See Editing & canvas tools.
• Screenshots and images: ⌘⇧I captures the preview for the agent, pasting an image over the dock saves it and stages its path, and draw-tool annotations arrive the same way.
• Files: drop anything from disk onto the terminal and its path is typed in, ready to reference.

The quick-edit pill (Select tool) is the shortcut path: small asks go from click to running agent without opening the dock at all.

Agents can drive Radial too

Every terminal inside Radial has the radial command on its PATH, and each session is told what it does: open any project by name or link, read a site's context, take a screenshot, scaffold a site, change text or a head tag in place, publish, and start a visible sub-agent in another site's terminal with radial agent start. Outside Radial the same verbs are available to Claude Desktop, Cursor, and Codex over MCP. See Command line and MCP.

The Home screen has its own agent: a terminal in your projects folder that starts when you first type to it. It is the one session to drive when you are away from the desk. Turn on Claude Code's remote control there and the Claude mobile app can open any site, pull its context, edit it, or start a visible sub-agent in that site's own terminal and watch it in the app.

Bring your own login

Radial launches the CLIs you already have, under the account you're already signed into: an existing claude login, an ANTHROPIC_API_KEY, or your Codex/Cursor auth. If a CLI is missing or signed out, Settings → Agent shows exactly what's needed with a one-click install or sign-in. For how this squares with each provider's terms, see AI accounts & terms.

Project rules

Put an AGENTS.md at the repo root and every agent reads it: brand voice, things never to touch, how to structure sections. Radial also briefs agents on its own conventions, like using the already-running dev server instead of starting their own, and (if .radial/figma.yml lists your design files) which Figma sources to pull from.