A clearinghouse for data,
built for agents.

How it works.

fivecell sits between data sellers and data buyers. We make sure the format is the same on both sides, the deal is settled, and there is a record of who got what.

Sellers list datasets and set the terms. Buyers find them, agree to the terms, and take delivery. Some give data away for research or consortia.

The bytes live wherever the seller chooses — on their own machine, or on ours. Either way, we put our name on the deal when the listing goes up. Buyers can check that signature before they touch a single byte.

Founding seller: Global Financials. Two more are signing up. A handful of buyers are coming online.

The format is the API.

You don't need a special client. Standard tools work: curl, any HTTP library, any code that speaks the web. Ask a small question and the answer comes back as plain JSON, parsed in two lines of any language. Pull a whole table and it streams as Apache Arrow or Parquet — the open standards — straight into Snowflake, BigQuery, Databricks, or DuckDB. Same columns, same types, same values either way. The choice is purely about size.

Sellers publish what each column means before they list. If the data doesn't match what was published, it doesn't ship. And every answer carries our signature, so you can prove what you received and who sent it. The recipes show how to check it; the machine-readable spec generates a typed client in your language.

How agents use it.

Most data marketplaces still require a human sales conversation — email, NDAs, custom connectors, manual delivery. An agent can't do any of that. Every step in fivecell's loop is plain REST and JSON, so an agent can:

• Discover datasets via GET /v1/inventory — public, paginated, no token needed. Each entry carries the seller's published column schema (name, dtype, nullable) and an acquire block telling the agent the exact HTTP call to make.
• Read the listing's machine-readable terms: licence, allowed operations, export caps, embargo dates, pricing.
• Acquire by following the acquire block — method, URL, headers, body template, required scope. No out-of-band knowledge needed.
• Ingest Arrow, Parquet, or JSON directly into the agent's tooling — bulk data in the typed columnar formats, small results and metadata as signed JSON.
• Verify the signature on every transfer and pin it in the agent's reasoning trace.

The agent does the loop on its own. The human who runs the agent gets the bill, the receipts, and a record of every dataset the agent touched. An agent that doesn't know our URL conventions starts at /.well-known/fivecell.json, or pulls the full OpenAPI 3.1 spec from /openapi.json to autoconfigure its client.

An agent that doesn't find the data it needs can POST /v1/feedback with no token — a free-form data request addressed to the operator or to a specific seller. Sellers subscribe to feedback_filed events to see what buyers are asking for; prospective sellers browse GET /v1/feedback publicly to decide what to list next.

Why we exist.

Every data deal today is its own project. A bespoke API. A bespoke SFTP drop. A schema you learn by guessing. Weeks of legal per buyer. The buyer writes parsers; the seller runs a portal team. Both sides know it should be standard. No one has the standing to make it standard.

That is our job. List once. Trade with anyone who shows up. Settle in whichever mode fits. Ship in formats we check on the way out. The wrapper is ours. The data is yours.