AI Scraper
The AI Scraper turns a sentence — “collect every product of a category, with its name, price and whether it is in stock” — into a program. AI is used once, to explore the site and understand what you meant; what it hands you is ordinary Python you can read, edit and publish, and every run after that uses no model at all.
- Create a scraper from one sentence
- Find your way around the Scraper Studio
- Test the program and publish it
- Know where AI is used — and where it is not
A sentence becomes a program
A new scraper is just a chat. You write what you want; the scraper asks one question if it needs an example value, explores the site, plans the steps, walks them in a real browser and prints the program that repeats that walk. The Studio then assembles itself around the same chat.
Create a scraper
- Open AI Scraper and press New scraper
In the project sidebar open AI Scraper. The empty page says Describe a scraper in your own words.
- Describe it
Say which site to start from, what to click or fill in, and what data you want back — no steps, no selectors. Name the inputs a run will change, such as a category or a search term.
- Answer the example question
If an input has no example, it asks for one — Which category should I use as the example? — because a site can only be explored with a real value.
- Watch it build
The plan runs at once. Pictures of every step arrive while it walks, and the rows appear when the program has run. See How a scraper is built.
The Studio
Once a plan exists, the page has two columns: the work on the left and, on the right, the Assistant and Execution history tabs. The topbar carries Test run and Publish.
| Part | What it shows |
|---|---|
| Preview · Validation | The work view, and the checks of the last build. |
| Browser playback | A full-page picture of each step, with a zoom bar. |
| Program session | The step log beside the picture: what each step did and how long it took. |
| Extracted results | The rows, with CSV export. |
| Scraper code | The program, always editable. |
| Logs · Site analysis · Inputs | Run details; how the site was read (plain requests or a browser, pagination, fields); the values the next Test run uses. |
Change it by asking
Ask the assistant for a change — “also read the rating”, “only items in stock” — and it modifies the existing program; the chat shows the change as a diff. You can also edit the code yourself. Either way you get a new version; press Test run to prove it.
Use it, and its limits
- A published scraper runs as a flow step or from a script, on a schedule or behind an API endpoint.
- It can follow a whole site: see Crawling a site.
- A search with no result ends with no rows and “Nothing found”, which is not an error — see Nothing found, proxies and captchas.