AI Connections
AI-Started Extractions
Ask Claude or any MCP agent to scrape product pages, build lead lists with emails and social links, or extract map place data — with credit estimates before anything runs.
With a full-access connection, you can ask your AI assistant to start a new extraction for you. These are zero-configuration extractions — no element picking, no selectors — the same technology behind the extension's automatic extraction modes, running in cloud browsers.
What agents can extract
| Tool | What you get |
|---|---|
scrape_pages | One row per URL from similar pages of a single site — products, listings, articles, job posts, real-estate pages. Structured data is captured automatically, no selectors — the usual answer for "get me data from these pages". Optional extras in the same run: page metadata (title, description, social-preview tags as fixed columns) and phone numbers |
extract_contacts | One row per site across many different sites — lead lists. Emails by default, social links on request (one column per platform), combinable in a single run — the AI-run counterparts of the Email Extractor and Social Link Extractor. An optional deep scan follows each site's internal links one level deep, so contact and team pages get covered too |
extract_google_maps | One row per Google Maps place page — name, rating, review count, address, phone, website, hours. Search and results links are rejected at estimate time with guidance; place links typically come from the extension's List Extractor |
Metadata always arrives as the same fixed columns for every URL — to pick individual head fields instead, build the step in the extension's Page Extractor. (Extractions saved before these tools existed keep running through the legacy create_extraction tool.)
Extractions run from a rotating scraping location by default — there's nothing to configure. When a site serves region-specific content (language, currency, availability), the agent can pin a country for the run instead.
Where the URLs come from
- Straight from the chat — paste up to 500 URLs into the conversation.
- From an existing table — point the agent at a link column of a table you've already scraped. This is the path for big jobs — up to 10,000 URLs for contact extractions, up to 4,000 for page scrapes and map place extractions — with no pasting. An optional row filter narrows the column to matching rows first (the same conditions as querying a table), and it's also how chaining works: scrape a list once, then have the agent visit every link in it.
- Discovered from the site — no URL list at all. Ask for a whole site or a section ("scrape all the products", "just the winter collection") and the agent maps the site's URLs, shows you the groups it found, and extracts from your selection. A fresh sitemap scan is quoted first — 25 credits, charged only when the scan actually finds URLs — and runs once you confirm; a recent scan of the same site is reused free, no confirmation needed.
Shopify stores get a fast path
When the site is a Shopify store, the agent can detect that and extract the whole catalogue — or chosen collections — as one clean row per product variant, with a credit estimate up front. The resulting table can even be exported as a re-import-ready Shopify product CSV.
Credit safety: estimate first, run on confirmation
An agent-started extraction never begins silently. The first request returns a credit estimate — "1,240 URLs ≈ 1,240 credits, you have 8,400" — so the agent can check with you before committing; the extraction only starts on a repeat call that carries your confirmation. For page scrapes the estimate also includes a free peek at the columns the first page yields (when that page is readable), so you usually see what you'll get before spending anything. Extractions cost about 1 credit per page, the same as any cloud run (see Credits & plans), and runs are refused when your balance doesn't cover the estimate. Credits can never go negative.
URL discovery follows the same pattern: a fresh sitemap scan quotes its fixed cost up front, is charged only when the scan actually finds URLs, and runs on confirmation — see the tool reference.
Scheduling
Agent-created extractions can carry a recurring schedule, like any other cloud task — so "check these 200 product pages every morning" is a single instruction. Automations you already have are manageable in the same conversation too: an agent can rename one, set or change its schedule, or pause and resume it.
Guardrails
The same rules as the rest of the product apply:
- LinkedIn is refused — it can't be used as a scrape target in any tool.
- Page scrapes work one site at a time —
scrape_pagestakes similar pages of a single site. Spanning many different sites is whatextract_contactsis for: one row per site, as many sites as your list holds.
Runs are asynchronous — the agent gets a run ID immediately and checks progress rather than blocking your conversation, with a live-progress link you can open in the dashboard.
Example conversation
You: Take the "Listing URL" column from my apartments table and scrape every page into one row each.
Agent: I asked
scrape_pagesfor an estimate: 486 URLs ≈ 486 credits (you have 9,200). The first page already yields these columns — name, price, address, floor area, year built. Run it?You: Yes.
Agent: Confirmed and started. I'll check progress… Done — 486 rows saved as a new table. Want it in Google Sheets?
You: Now find contact emails for each agency in the "Agency site" column.
Agent: That's 112 different sites —
extract_contactsestimates about 112 credits, one row per site. Run it?
Related
MCP Tool Reference
Every MCP tool a connected AI agent can use in Ultimate Web Scraper — querying tables, cleanup, exports, automations, site discovery, and Shopify extraction.
AI Connection Security
How Ultimate Web Scraper keeps AI agent access safe — workspace-scoped connections, read-only mode, instant revocation, rate limits, and credit protection.