Open the annotation browser (headed Chrome, persistent profile) and optionally navigate to a URL. The comment/pencil toolbar is injected into every page automatically. Call this first; the human then browses and marks up pages themselves. ALWAYS follow this with annotate_wait in the same turn — the human expects pressing Send to reach you, and it only does while something is parked.

Park until the human presses Send in the review panel, then return that batch. This is the main hand-off and the ONLY way a send reaches you on its own — an MCP server cannot start a turn, so with nobody parked a send just sits in the store until someone asks for it. Call this immediately after annotate_open and again after handling each batch. It is cheap to park: some clients (Claude Code) move a call still running after two minutes to a background task, so you keep working and the batch arrives as a task notification. The overlay shows the human a live indicator of whether anyone is parked.

List annotation marks as a readable briefing, grouped by page. Use this to check what is pending without waiting for a send.

Re-read the most recent Send batch without waiting. Useful if the conversation moved on and you need the original request again.

Screenshot the current page with the drawings visible but the toolbar hidden. Pass a markId to crop tightly around that mark — the fastest way to see what a squiggle actually meant.

Mark items as done once you have acted on them. They stay visible on the page but dim, so the human can see what you handled.

Delete marks permanently. With no ids, deletes everything.

Close the annotation browser. Marks are kept on disk.