Data Marketplace for Claude

What you can do from Claude

The buyer's journey, and the seller's. Each stage is a tool Claude calls for you.

1. discover

   Find products that might answer the question.

   search_datasets

2. evaluate

   Read one product's spec, licence, data dictionary, quality and per-file hashes before committing to it.

   describe_product

3. sample

   Look at rows: a seller-permitted sample, or a whitelisted equality query behind a licence.

   describe_datasetquery_dataset

4. compute

   Get an aggregate over a licensed dataset without ever holding the rows.

   compute_dataset

5. acquire

   Buy the product with the calling key. Final once delivered.

   acquire_product

6. use

   Take delivery and verify the bytes against the hashes published at the evaluate step.

   list_purchasesget_delivery

7. bounty

   Where the buyer's journey goes when the catalog has nothing suitable: say what is missing, and let a provider supply it.

   search_bountiesdescribe_bountycreate_bountylist_own_bounties

8. publish

   The seller's journey: create a draft, upload files, wait for the scans, document, price, publish.

   create_productupdate_productpresign_uploadfinalize_uploadget_scan_statusset_documentationset_sampling_policyset_pricingpublish_productlist_own_products

Prompts you can paste

Say these to Claude once the command above has run.

• “Find me EV charging utilization data from the last 3 years.”

  Claude calls search_datasets and comes back with candidate products, their coverage and their prices — the catalog is searchable with a key, the rows behind it are not.

• “Compare the available datasets and choose one with strong provenance and clear commercial rights.”

  describe_product returns each candidate's quality breakdown, licence class and per-file sha256 commitments, so the comparison runs on measured values rather than on the seller's description.

• “Don't expose raw user records — use privacy-preserving compute to get the aggregate.”

  compute_dataset returns statistics over a licensed dataset without ever exposing rows, and suppresses small groups. Until a licence is assigned it answers 402 no_license — that refusal is the honest result, not a failure of the example.

• “If nothing satisfies these requirements, create a bounty describing what I need.”

  search_bounties looks for an open request first; create_bounty posts one with explicit acceptance criteria. Posting costs nothing, and providers fulfil bounties in the web app.

Prefer compute_dataset to acquire_product when an aggregate answers the question: it returns statistics over a licensed dataset without exposing raw records, so a hypothesis costs no budget.

When no product matches, check for an open bounty with search_bounties; if none matches, post one with create_bounty and explicit acceptance criteria — required columns, coverage window, geography, cadence, licence class, PII posture. Posting does not charge you; providers fulfil bounties in the web app.

Why the results are trustworthy

• Quality: describe_product returns a quality breakdown, so candidates rank on a measured score rather than on the seller's description.
• Licence: the same spec carries the licence class and machine-readable open-data terms, so the licence check precedes the purchase.
• Provenance: a purchasable product pins a current version whose per-file sha256 commitments (version.files[].sha256) are published before purchase.
• Integrity: get_delivery returns that commitment again as commitmentHash beside the presigned links, so delivered bytes can be hashed against what was advertised.
• Personal data: sampling availability carries inspectedScope — what the scan actually inspected, null when unrecorded — so a clean scan differs from an absent one. Sellers read piiStatus, piiFlagged and piiReviewApproved for their own files via get_scan_status.

What to expect when you buy

• A priced purchase needs a person: acquire_product answers 402 payment_method_required with a checkout link.
• Purchase-scoped keys carry three ceilings — per transaction, per 24 hours, per 30 days, $20 each by default — and answer 402 spend_limit_exceeded past one.
• A human confirms every priced purchase through the checkout link. Claude can reach the point of sale; it cannot spend money without you.
• An acquisition is irreversible. Once the files are delivered the purchase is final and cannot be un-delivered — check the licence and the quality breakdown before the buy step, not after.