Skip to main content
Every market runs a Model Context Protocol server. Point an MCP client at it and an assistant can query company data directly, without you writing request code.

​
Endpoints

One server per market, each authenticated with the same API key you use for the REST endpoints. See Authentication.
Connect more than one server when you want an assistant to compare markets. Each server only knows about its own country, so comparing the US with Germany means connecting both.

​
Connecting a client

Most clients take a URL and a header. For Claude Code:
Clients that use a JSON configuration file usually take this shape:
Name each server after its market. An assistant with tdc-us and tdc-de connected can tell them apart; two servers both called tdc cannot be reasoned about.

​
Tools

Each server exposes six tools. They are a deliberate subset of the REST API, not a mechanical translation of all eleven endpoints — batch and classification endpoints are left out because they serve bulk pipelines rather than conversational use. list_filters is described to the assistant as the tool to call first. The same rule that governs the REST API applies here: valid values cannot be guessed, and an unrecognised one returns an empty result rather than an error. See Errors.

​
What you can ask

The tools are designed to be composed, so useful questions are the ones that need several steps:
  • “How many industrial biotech companies are there in the US compared with Germany?”
  • “Find companies in Bavaria with a website and turnover above €10m, then show me the group structure of the largest three.”
  • “What filters can I use for France, and which of them do not exist for the US?”
The assistant calls list_filters to discover the right codes, then filter_companies to run the query. You do not need to know the classification scheme before asking.

​
Rate limits

MCP traffic is limited to 300 requests per minute per key, higher than the 60 per minute that applies to the REST endpoints, because a single question can trigger several tool calls.
A throttled tool call comes back as a successful MCP result whose content is a 429 document. The assistant may read that as data and answer from it rather than reporting a failure. If answers suddenly become vague or contradict earlier ones, check whether you are hitting the limit.

​
Limits worth knowing

  • The server is read-only. No tool writes, updates or deletes anything.
  • A key reaches all four markets, so connecting a market server is a choice about what the assistant should see, not a permission boundary.
  • Tool results count against your quota the same way REST calls do. An assistant exploring a broad question can spend a lot of calls quickly.
Tell us which questions your assistant handles badly at support@thedatacity.com. The tool set is curated by hand, so what agents actually struggle with is what we change next.
Last modified on July 30, 2026