The read-plan-approve-execute loop
First the agent explores: what files exist, what is relevant. Then it plans concrete actions. You approve. Then it executes and reports. This loop is what makes agents safe enough to trust with real work.
What data leaves your machine
Only the task context — the files and commands relevant to what you asked — is sent to the AI service. Everything else stays local. For privacy-sensitive work, desktop clients like MOOGH keep processing on your side wherever possible.
Reviewing what changed
After execution, the agent summarizes the changes and the results. You can ask it to adjust, revert specific actions, or continue — the history is always available.