/datasets

name: datasets
description: 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). Also when the user holds a dataset_id that turns out to be a marketplace row, or asks what a ready dataset costs. Not for fresh, specific, or recurring collection, which is scraping and belongs to the scrape skill.

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.