Agent Mode

Connect agent CLIs

Claude, Codex, Cursor, Antigravity and Copilot. Connected ones plan and run tasks. A missing CLI can be installed from Settings, then connected.

Sign in on an agent’s row runs that CLI’s own sign-in inside the app. If a connected agent needs sign-in or stops answering, the workspace warns you before a prompt fails.

Agent CLIs run under your own accounts and licences. What they send, retain and charge for is governed by that vendor, not by OpenPrompt.

Run a task

Describe the work, or @mention a project agent such as @Frontend.

• Plan answers without editing files.
• Build can change them.
• Smart Enhance rewrites your draft with this project’s instructions.
• Quick Enhance only fixes the wording.

Named project agents

Give a project named agents (such as @Frontend) and address them by name. The Free plan includes three named agents per project; Individual and Team have no limit.

Routed pipelines

A connected CLI plans the task, then each sub-task runs on an agent. Expand a step to read what it changed, and stop a run from the composer while it is still going.

The cost and task router reads each prompt and sends easy work to cheaper models, keeping premium ones for hard problems.

The Workbench

The OpenPrompt chat and your agent CLIs share one row of tabs. Nothing is open at first: open the chat, or Claude Code or any other installed CLI on its own tab. A CLI stays running while you move between Captures, Git and Usage, and closing the chat tab keeps the conversation.

Before you let an agent build

• Agents edit real files. A coding agent creates, changes and deletes files in the folders you grant it, and runs commands on your machine. Commit before you start, so anything unwanted can be undone.
• AI output can be wrong. Review generated code before you rely on it.