MCP tool reference

Read tools include list_sites, get_analytics, get_timeseries, get_breakdown, get_realtime, get_revenue, list_experiments, get_experiment, get_experiment_results, list_goals, list_funnels, get_funnel, list_events, list_personalizations, get_personalization, list_annotations, get_sites_overview, get_signups, get_crawlers, get_visitor, get_site_settings, check_install, get_notification_prefs, and suggest_tests.

Write tools include create_site, create_experiment, update_experiment, archive_experiment, add_variant, update_variant, set_variant_weights, delete_variant, launch_experiment, pause_experiment, stop_experiment, declare_winner, create_goal, create_site_goal, update_site_goal, delete_site_goal, create_funnel, update_funnel, delete_funnel, create_personalization, save_personalization_draft, publish_personalization, pause_personalization, restore_personalization, delete_personalization, create_annotation, delete_annotation, update_site_settings, update_notification_prefs, and set_share_link.

For analytics, get_timeseries shows trends over time, get_breakdown ranks one dimension (page, entry_page, exit_page, channel, source, referrer, utm_campaign, country, region, city, browser, os or device, up to 50 rows), and get_realtime shows who is on your sites right now. get_analytics, get_timeseries and get_breakdown accept filters such as country, source, device or page. get_analytics also returns the previous period for comparison. get_sites_overview compares all your sites at once, get_signups shows where signups came from (no emails), get_crawlers shows AI and search-engine crawler visits, and get_visitor shows one visitor's journey. create_annotation marks a deploy or campaign on the analytics timeline.

For setup, get_site_settings and update_site_settings read and change a site's privacy, timezone, currency and signups email settings (secrets are never returned), check_install tells you whether the snippet is installed and sending data, get_notification_prefs and update_notification_prefs manage your own email notifications, and set_share_link turns the public share link on or off. Turning it on again keeps the same link unless you ask to rotate it. suggest_tests asks AI for A/B test ideas for one page of your site and needs the Growth plan.

Write tools need write access. An OAuth connection is read-only unless you choose Read and write when you connect, and a read-only API key can't use them either.

create_goal adds a goal to an experiment. For a site-wide goal (Analytics → Goals), agents use create_site_goal. Experiment goals can't be changed or deleted over MCP. The site's signup goal, the one the New signups report uses, can't be deleted or re-pointed over MCP; do that in the dashboard.

add_variant and update_variant define what a variant does: DOM changes for a visual experiment, a redirect URL for a redirect experiment. set_variant_weights sets the traffic split across all variants and must sum to 100. delete_variant also deletes that variant's events, and is refused while the experiment is running. update_experiment can also set which pages an experiment runs on and its optimization mode.

Each tool tells your AI assistant whether it only reads, and whether it is destructive. The destructive tools are delete_variant, delete_site_goal, delete_funnel, delete_personalization, delete_annotation, stop_experiment, declare_winner, archive_experiment, and set_share_link: they delete something, end an experiment, or publish or break a share link in a way you can't undo. Many assistants ask you to confirm these before they run.

Tool availability still respects tenant isolation, site scope, key permissions, rate limits, and plan status. Review any proposed write before approval, especially launch, stop, and winner decisions.

Prompts

PageDuel also offers three ready-made prompts. In Claude.ai or ChatGPT you can pick one instead of writing the instructions yourself:

• Weekly review (weekly_review): the last 7 days for one site, with traffic, sources, pages, signups and running experiments, ending with what changed and one next step.
• Diagnose a traffic drop (diagnose_traffic_drop): finds which channel, source, country, device or page fell, checks your annotations, the snippet install and crawlers. Give it the day the drop started, or leave it out to compare the last 7 days with the 7 days before.
• Plan an experiment (plan_experiment): for one page path, looks at the baseline, asks AI for test ideas, and sets up a draft experiment with its variant and primary goal. It never launches anything: it asks you to confirm first.