dbt + LLM Analytics

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

Landing page  |  Try the Lab

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:

1. What is dbt?
2. Setup
3. Staging models
4. Mart models
5. Testing and documentation
6. LLM analytics

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.