poke-browser

A Chrome extension + MCP server for natural-language browser automation

Built by Poke — the AI assistant from interaction.co — together with Leo Kök.

poke-browser on GitHub

• `navigate_to` — Open URLs and wait for load completion on the chosen tab.
• `find_element` — Locate elements by CSS, text, ARIA, or XPath; queries respect open shadow roots (and the same-document tree you’d expect for complex widgets and “portal-style” UI mounted in the page).
• `click_element` — Click by selector or viewport coordinates; always hovers ~1s before the click so hover menus and delayed affordances can appear.
• `type_text` — Type into inputs and contenteditable regions; optional `clear` (default true) wipes existing text before typing, or set clear: false to append.
• `get_dom_snapshot` — Compact DOM tree with tags, roles, labels, bounds, and interactivity hints.
• `capture_and_upload_screenshot` — Capture the visible tab and POST it to your upload endpoint (or fall back to inline base64 when upload isn’t configured).
• `get_accessibility_tree` — Semantic nodes in reading order for screen-reader–style reasoning.
• `scroll_window` — Scroll by position, delta, direction, or “scroll into view” for a selector.
• `managetabs` — List, open, close, and switch tabs in the connected Chrome profile.
• `browser_guide` — In-repo Markdown playbook: every tool, common flows, and troubleshooting.

After every `click_element` call, inspect the page again with `get_dom_snapshot` (or related tools) before the next action. Clicks often open modals, slide-overs, or rerendered regions; a fresh snapshot keeps the model aligned with what the user actually sees.

1. MCP server (npm)

Run the published launcher (recommended):