[CAIS’26 Demo] AgentClick: A Human-in-the-Loop Review UI for Autonomous Agents

AgentClick is a skill-based plugin that adds a browser-based review layer on top of terminal agents. Instead of agents going straight from decision to action, AgentClick inserts a review step — the agent proposes an action, a purpose-built browser UI opens for the user to inspect and edit, and then the agent continues with the user’s approved changes. It supports email review with preference learning, plan review as a DAG, GitHub-style code review, trajectory inspection, action approval with risk levels, and memory management. AgentClick is framework-agnostic and provides explicit support for Claude Code, Codex, and OpenClaw.

This paper is accepted as a demo at ACM CAIS 2026. Project website: AgentClick. Code is available on GitHub.