r3
A tool for reviewing AI-generated HTML pages, markdown documents and code diffs by leaving comments directly on the artifact, then sending that feedback back into the originating coding agent to act on.
r3 lets a person open an interactive HTML page, a rendered markdown document, or a code diff produced by an AI coding agent, select a specific element or passage, and leave a comment that continues the conversation with that agent rather than describing the problem in a separate chat. Agents publish artifacts to r3 through a CLI or an HTTP API, so any agent able to run the CLI can use it, and a bundled browser demo with a scripted agent lets a user try HTML, markdown and diff feedback without a live agent connected.
A guide command explains how an agent should publish an artifact, listen for feedback and reply to it; local Claude Code and Codex publications wire up feedback delivery automatically, while other setups can copy a feedback-fetch command into the agent's own session. A fetch only acknowledges feedback once it has been printed successfully, so a failed read leaves the feedback pending rather than silently dropping it, and a history flag can read past feedback without acknowledging it or registering as a listener. It installs globally with npm or bun, or runs directly with npx.