AIData.Broker

Measured conclusion: Deep-tested. L2 outcome: tools_ok.

Published prices, market rates, buyers, deals and rulings for AI training data, and valuations.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "AIData.Broker": {
      "url": "https://aidata.broker/mcp",
      "transport": "streamable-http"
    }
  }
}

Measured tools (10)

• get_market_rates read-only

  Median, middle half and range per comparable unit (per hour of video or audio, per image, per task, per company, per title, and more) for each kind of AI training data and each side of the market, computed from published prices. Without `kind`, returns every rate the evidence supports a median for.

  Parameters:

  • kind — string — The kind of data: company (Company records), code (Source code), video (Video and footage), audio (Audio and speech), images (Images), text (Text and archives), environment (Datasets, tasks and environments), physical (Physical-world data), people (Work by people), products (Data products).
  • side — string — The side of the market: pay (Paid to the people who make it), offer (Offered to holders), licence (Sales that happened), listing (Asking prices off the shelf), collection (Made to order), cost (Cost to produce).

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• search_prices read-only

  The individual price observations behind the market rates: what a lab, vendor, marketplace or court said data was paid, offered, listed or sold for, with the figure as published, the figure per unit, the date, the source's own words and a link. Newest first.

  Parameters:

  • kind — string — The kind of data: company (Company records), code (Source code), video (Video and footage), audio (Audio and speech), images (Images), text (Text and archives), environment (Datasets, tasks and environments), physical (Physical-world data), people (Work by people), products (Data products).
  • side — string — The side of the market: pay (Paid to the people who make it), offer (Offered to holders), licence (Sales that happened), listing (Asking prices off the shelf), collection (Made to order), cost (Cost to produce).
  • limit — integer — How many rows to return (default 50, at most 50).
  • query — string — Words that must all appear, such as a company ("Shutterstock"), a language ("hindi") or a kind of record ("egocentric").

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• value_data read-only

  What published prices imply for a holding of data: a low-to-high range in US dollars with the arithmetic behind it. Call it first with only `kind` to get the questions and their answer ids, ask the holder, then call again with `answers`. Questions not answered are taken at their defaults and marked as assumed. An indication, not an offer.

  Parameters:

  • kind (required) — string — What is being valued: records (Company records), code (Codebases), tasks (Training tasks), video (Video and footage), speech (Audio and speech), images (Images), text (Text and archives), environments (Simulated environments), robotics (Robot and physical-world data).
  • answers — object — Answer ids by question id, as returned in `questions` (for video, for example {"hours": "1000-10000"}); every answer is one of the listed answer ids, sent as text.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• find_buyers read-only

  Organisations with a public way to sell them data, with their own published terms: what they want, what they pay, when they pay, who removes personal data, eligibility, and what they pay an introducer. Largest published figure first.

  Parameters:

  • kind — string — The kind of data: company-records (Company records), source-code (Source code), video (Video and footage), audio-speech (Audio and speech), images (Images), text-archives (Text and archives), datasets-environments (Datasets, tasks and environments), physical-world (Physical-world data), expert-work (Expert and human work).
  • limit — integer — How many rows to return (default 10, at most 50).
  • query — string — Words that must all appear in the buyer's name or terms.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• find_vendors read-only

  Vendors that sell training data, RL environments, evaluations, expert work or annotation to AI developers, each with what it says it sells and where its data comes from in its own words, and whether it says it pays holders for data. A to Z, not ranked.

  Parameters:

  • buys — boolean — Only vendors that say they buy, license or pay for data from its holders.
  • from — string — Where it says its data comes from: licensed, customer, experts, captured, synthetic, replica, public ("licensed" means licensed from the companies or creators that hold it).
  • limit — integer — How many rows to return (default 10, at most 50).
  • query — string — Words that must all appear in the vendor's name or what it sells.
  • sells — string — What the vendor sells: rl-environments, datasets, human-data, evals, annotation, simulation.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• find_lots read-only

  Every lot publicly on offer in the catalogues the lot index reads each day (companies' records, codebases, footage, speech, environments), in one vocabulary: kind, industry, years of history, systems, size, what its catalogue says about the rights and whether it is ready, each linked to its catalogue's own page. We have not checked any catalogue's claims. No holder is named. Longest history first.

  Parameters:

  • kind — string — The kind of data: company-records (Company records), source-code (Source code), video (Video and footage), audio-speech (Audio and speech), images (Images), text-archives (Text and archives), datasets-environments (Datasets, tasks and environments), physical-world (Physical-world data), expert-work (Expert and human work).
  • limit — integer — How many rows to return (default 10, at most 50).
  • query — string — Words that must all appear in the lot's title, line of business or systems.
  • rights — string — "settled": the lot's own catalogue says its holder has the rights with nothing left to confirm; "open": the catalogue names conditions or consent still to confirm. We have not checked either.
  • industry — string — The holder's industry, as this site names industries.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• search_deals read-only

  Licences of data to AI companies disclosed since July 2023: licensor, licensee, content, use (training or display), term, value where disclosed, with a verbatim quote and the source. Newest first.

  Parameters:

  • use — string — What the data is used for.
  • limit — integer — How many rows to return (default 10, at most 50).
  • query — string — Words that must all appear, such as a licensor, a licensee ("OpenAI") or a kind of content ("news").
  • since — string — Only deals on or after this date (YYYY, YYYY-MM or YYYY-MM-DD).

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• search_cases read-only

  Lawsuits, rulings, settlements and regulatory actions over data used to train AI, with parties, claim, amounts, outcome, status and the source. Newest first.

  Parameters:

  • kind — string — The kind of case: guidance, lawsuit, regulatory action, ruling, settlement.
  • limit — integer — How many rows to return (default 10, at most 50).
  • query — string — Words that must all appear, such as a party ("Anthropic"), a court or a kind of content ("books").

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• check_software_terms read-only

  For a business system (Slack, Microsoft 365, Google Workspace, Salesforce, HubSpot, Zendesk, Zoom, GitHub, Notion and more): whether the holder can export its data, what the vendor's API terms say about training AI on data pulled through them, our reading of the content's rights risk, and the clauses word for word. Research from public sources, not legal advice. Without `system`, lists the systems on file.

  Parameters:

  • system — string — The system's name, such as "Slack" or "Zendesk".

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• list_datasets read-only

  Every table behind these tools as a CSV and JSON download (our compilation of the tables under CC BY 4.0; quotes keep their sources' terms, and article text is not covered), with row counts, review dates and the permanent copies with their DOIs.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

Publisher website

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.