prumi

The Prumi server and the prumi CLI. AI agents connect to its MCP endpoint and build dashboards on a company's official metrics; Prumi checks every page, hosts it, and lets the data team certify it.

Import name: prumi. It holds the catalog, the query engine, the BI validator and the registry, and defines the adapter protocol that source adapters implement. Prumi reads metrics through an adapter package; the first one is prumi-adapter-dbt.

• Agents connect over MCP (Streamable HTTP, /mcp), find the official metrics, learn the shape of the data (never its values) and publish a dashboard: a spec plus any HTML.
• Numbers come only from the source, computed at the grain they're shown. Validation renders each page in headless Chromium and checks that every number it shows is a value Prumi delivered.
• Dashboards run in a sandboxed frame inside Prumi's own page, which shows the trust badge and data freshness. The data team certifies a version from the CLI; any later change makes it a draft again.

Install

With uv, Prumi and its dbt adapter:

uv tool install "prumi[dbt]"
prumi install-chromium
prumi serve --config prumi.toml

pip install "prumi[dbt]" installs the same packages. prumi.toml points at your dbt project's semantic manifest and warehouse: see Use it with your metrics and every setting.

Requirements: Python 3.12 or later, and headless Chromium, which prumi install-chromium installs with Playwright (add --with-deps on Linux if system libraries are missing); the server won't start without it.

The quick start on a demo project, configuration and where data goes are in the Prumi README; the BI rules Prumi checks are in docs/bi-rules.md.