dbt + LLM Analytics

Learn dbt on real data, then give it a natural-language interface with Claude.

This is a hands-on tutorial. You build a tested dbt project on real National Park Service biodiversity data (5,824 species, 23,296 observations across 4 national parks), then connect it to Claude so you can ask questions in plain English and get real answers back.

What you'll build

A clean data pipeline:

seeds (raw CSVs)  ->  staging (cleaned)  ->  marts (ready to analyze)

Then an MCP server that lets Claude query it safely:

You: Which park has the most endangered species sightings?

Claude: (lists the tables, reads the schema, writes the SQL, runs it) Yellowstone, with the highest at-risk observation counts of the four parks...

Claude can read everything and change nothing — queries are locked to read-only SELECTs.

Try the lab

The interactive lab runs in your browser — no local setup required:

Start the Lab on Killercoda

Run it locally

Requires Python 3.11+.

pip install -r llm_query/requirements.txt
cd dbt_project
dbt build --profiles-dir .

Then open the repo in Claude Code (the MCP server connects automatically via .claude/settings.json) or wire it into Claude Desktop using llm_query/claude_desktop_config.example.json.

The tutorial

Work through these in order:

How it fits together

Piece	What it does
`dbt_project/seeds/`	Raw NPS biodiversity CSVs
`dbt_project/models/staging/`	One cleaned view per source
`dbt_project/models/marts/`	Three analysis tables with tests + column docs
`llm_query/mcp_server.py`	MCP server: `list_tables`, `describe_table`, `run_query` (SELECT-only)
`.claude/settings.json`	Auto-connects the MCP server in Claude Code

The database is DuckDB — it runs in-process, so there's no server to start and no credentials to manage.

Data source

National Park Service biodiversity dataset, covering Bryce, Great Smoky Mountains, Yellowstone, and Yosemite.

FilesExpand file tree

README.md

Latest commit

History