Working with the agent

Ask in plain words

Describe the data the way you would to a colleague:

500 customers from 20 countries as CSV.
A pytest test for validate_user on 50 different users.
SQL inserts for a clinic: departments, doctors in each department, and their appointments.
The same orders file, but 10,000 rows and 30% of them canceled.

Say what matters to you: the format, how many rows, which fields must agree with each other, the shares you need, the language or country. Anything you leave out, the agent decides and writes into the config, where you can see it.

What the agent does

1. Reads the TDC guide, once per session.
2. Looks up real-world values in the data packs — names, cities, IBANs. When a language or country is missing, it installs the pack; with the MCP server, your client asks you to allow that first.
3. Writes the .tdc config and checks it with the engine until no error is left.
4. Generates a preview of the first rows, or the whole file.
5. Summarizes the file — row count, the split of each column, the ranges — to show it holds what you asked for.
6. Hands you the file and the config.

With the skill alone, the agent does all of this with the tdcv2 command in its terminal. With the MCP server, it calls the server's tools instead.

The server's tools

Tool	What it does
tdc_read_docs	the guide, the traps check does not catch, test code in five languages, several linked tables in one file, any page here
tdc_find_packs	finds installed data packs by words ("last name", "iban") and gives their addresses
tdc_pack_add	installs packs for another language or country — tdcv2 pack add
tdc_check	every diagnostic, with code, line, column, hint and a "did you mean"
tdc_generate	a preview of up to 50 lines, or the whole file
tdc_peek	a short summary of a CSV, JSON or SQL file in the project; for a JSON object of several tables, each table, and how many <x>_id values have no match in <x>s.id
tdc_format	formats a config the way tdcv2 format does

The server also offers the guide as a prompt, tdcv2_guide, and four pages as resources: tdcv2://guide, tdcv2://traps, tdcv2://from-code and tdcv2://linked-tables.

Review what you get

The config is what you keep. It is a short file that states every rule the data follows, and reading it is quicker than reading the data — Your first dataset explains what you will see in one. Keep it in the project beside the file it made.

The same config with the same seed gives the same file, byte for byte. Rerun it yourself without the agent:

npx -y tdcv2@0.3.6 customers.tdc -o customers.csv

To change the data, change the config: ask the agent ("make it 10,000 rows, a third from Germany"), or edit the file and rerun it. tdcv2 check customers.tdc tells you whether an edit is valid before you generate.