Introduction
Metrics Book is Agent First BI: business intelligence that an AI agent builds, maintains and answers questions from. It is open source and runs on one virtual machine.
dlt loads your data, dbt transforms it, DuckDB stores it, and anfra holds the governed metrics. One API with an MCP connector sits in front, so a person using Claude, or any MCP client, gets the numbers the team agreed on.
How it works in one paragraph
Section titled “How it works in one paragraph”Everything that defines your setup is a file in git: pipelines, dbt models, metrics, terms, reports, access policy and budgets. An AI agent changes those files through pull requests, and a person reviews and merges them. The stack reads the files and serves the result through one API. People ask questions through Claude, and every query is checked against the access policy and written to an audit log.
Who it is for
Section titled “Who it is for”- Teams that want governed metrics in git. One definition per metric, with a description an agent can read, instead of the same number defined again in each dashboard.
- People who direct an agent and know the outcome, not dbt. The agent asks in business words (“Are cancelled orders an order?”) and takes the technical decisions to the repository’s standards.
- Teams that want Claude connected to their data. Access is decided per dataset, metric and report, and every query is audited.
One deployment serves one organisation. Row-level security, writing definitions through the API, and scheduled report delivery are not in version 1.