inspect_page returns text rather than an image, costs 1 API request, and is named as the call to make before run_sequence or record_video so the selectors going into them are real ones. A video costs 3 requests and every output inside a sequence costs 1, so a well-planned run and an exploratory one differ by a measurable amount rather than a feeling. Polling an async render is free of quota entirely, which is what makes the asynchronous video path practical. act_on_page is the outlier: it is the one tool where you author neither selectors nor steps, and what comes back is a trace of what it decided to do on your behalf.
A hosted browser capture and automation service covering 17 tools — screenshots, PDFs, social cards, video recording, page observation, goal-driven acting, visual diffing, persistent sessions and job polling — with each tool naming what it costs in API requests.
- take_screenshot with device emulation, ad, banner, chat and tracker blocking, full-page capture with scroll control, element selectors and clipping, geolocation and timezone, injected CSS or JS, dark mode, and macOS or Windows frames over gradient or glass backgrounds, returning PNG, JPEG or WebP.
- generate_pdf from a URL or HTML with margins, header and footer templates, page ranges, landscape and scaling, saved to disk with its path returned.
- create_og_image, which builds an Open Graph card from built-in templates or custom HTML with a title, subtitle, logo and colours.
- run_sequence, a multi-step automation in one browser session — navigate, click, fill, select, and capture screenshots, PDFs or diffs along the way — where each output counts as 1 API request.
- record_video, which turns a sequence into MP4, WebM or GIF with cursor highlighting, click effects, smooth movement, step notes, a browser frame and backgrounds; it costs 3 API requests and works best at 5-15 action steps.
- Seeing a page before acting on it, at two levels of detail: inspect_page, returning a structured map of interactive elements, headings, forms, links and images each with a unique CSS selector, and observe_page, returning a token-budgeted observation with id-indexed elements, a heuristic page-type classification — login, signup, search, article, form or generic — and grouped suggested actions.
- act_on_page, which takes a URL and a plain-English goal and runs an observe, plan, act and verify loop server-side until the goal is met, returning a structured trace of every action plus a success or failure status.
- import_agent_trace, which converts a page-agent or browser-use action trace into a re-runnable sequence, with a dry run that returns the translation without saving it.
- visual_diff, which compares two pages or HTML strings pixel by pixel and returns the diff image, the changed pixel count and the percentage changed.
- Persistent sessions: create_session keeps a live browser page with its cookies, localStorage and auth state for reuse across calls, with list_sessions for what is still alive and destroy_session to free one early.
- list_devices for the viewport presets such as iphone_14_pro and macbook_pro_14, check_usage for usage against plan limits, and list_jobs with get_job for async renders, both free of request quota.
PAGEBOLT_API_KEY. Enough request quota on the plan for what you are doing, since the tools price themselves in API requests and check_usage reports where you stand. A Starter plan or above for persistent sessions, which expire after 10 minutes of inactivity against a hard cap of 30 minutes.
One command plus a key — npx -y pagebolt-mcp, then supply credentials