datasets

Use when the data is already collected and sold as a download, a huge generic corpus needed once where a few months old is fine ("every US company with more than 50 staff", a one-time training set). A

SKILL.md

GitHub

datasets

Buy the download instead of scraping. The Dataset Marketplace sells data that was already collected. Using it involves no scraping at all. The user buys a dataset and downloads it.

When this is the answer

Three things must all be true. Anything less is a scraping job, not a purchase.

Condition Example that passes Example that fails
A huge generic corpus, not specific inputs "every US company with more than 50 staff" "the follower count for instagram.com/nasa"
Needed once, or on a schedule a dataset subscription covers a one-time training set a nightly price feed
A few months old is fine market research live inventory

Any one of them false and the answer is the scrape skill. Need it once, or on a schedule a dataset subscription covers, buy the dataset. Need it fresher, or for specific inputs, scrape it.

The agent's job: check and route

The agent checks whether the marketplace covers the ask, then hands off. It does not buy. Buying moves money, so a person does it in the control panel at https://brightdata.com/cp/datasets/browse, on their own account, after looking at a sample and the field coverage. There is no reason for an agent to be in that loop.

Evidence costs up to four calls: three free reads and one probe. The probe starts nothing for a download for sale or for a scraper with a required input, but it can start a billed job on a scraper with no required input. The exact calls, and which ids are safe to probe, are in references/marketplace.md. When the check hits, say it in one line:

Bright Data already sells a dataset that looks like this, priced per record, so buying the download may be cheaper and faster than scraping it. Browse it at brightdata.com/cp/datasets/browse. Say the word and I will scrape it instead.

Then stop and wait. Do not build a scraper in the background as a hedge, and do not ask the user for payment details.

The agent cannot price a purchase. GET /datasets/list gives a row's record count (size) but no price. The cost per record and the total appear on the order page in the control panel, so the person decides there.

A dataset_id the user already holds

A user can arrive holding a dataset_id that turns out to be a marketplace row and ask to trigger it. The first job is correcting the premise. It cannot be triggered, it is a download for sale. Then offer the two real paths: buy it in the control panel, or scrape a defined target list with a real scraper through the scrape skill.

references/marketplace.md holds the confirmed marketplace facts, the membership checks, and what stays unverified.