Metabase represents database metadata — synced databases, their tables, and their fields — as a tree of YAML files. Files are diff-friendly: numeric IDs are omitted entirely, and foreign keys use natural-key tuples like ["Sample Database", "PUBLIC", "ORDERS"] instead of database identifiers.

This repository contains the specification, examples, and a CLI that converts the table_metadata.json downloaded from a Metabase instance into YAML.

The format is defined in core-spec/v1/spec.md (v1.0.4). It covers entity keys, field types, folder structure, and the shape of each entity.

Reference output for the Sample Database lives in examples/v1/ — both the raw table_metadata.json and the extracted YAML tree.

Metadata is fetched from Metabase's GET /api/ee/serialization/metadata/export endpoint as a table_metadata.json file — a flat JSON document with three arrays (databases, tables, and fields) streamed so even warehouses with very large schemas can be exported without exhausting server memory.

The endpoint accepts three boolean query parameters that opt sections in or out — they all default to false, so requests must explicitly set the sections they want:

• with-databases — include the databases array.
• with-tables — include the tables array.
• with-fields — include the fields array.

A typical full export sets all three to true:

GET /api/ee/serialization/metadata/export?with-databases=true&with-tables=true&with-fields=true

The CLI turns that JSON into the human- and agent-friendly YAML tree described in the spec:

bunx @metabase/database-metadata extract-table-metadata <input-file> <output-folder>

• <input-file> — path to the table_metadata.json downloaded from Metabase.
• <output-folder> — destination directory. Database folders are created directly under it.

The bundled spec can be extracted to any file — convenient for agents that need to read it locally:

bunx @metabase/database-metadata extract-spec --file ./spec.md

Omit --file to write spec.md into the current directory.

The following is the default workflow for a project that wants to use Metabase metadata. It is a convention, not a requirement — teams are free to organize things differently.

Create a top-level .metadata/ directory and add it to .gitignore. This is where the raw table_metadata.json and the extracted databases/ YAML tree live:

.metadata/
├── table_metadata.json
└── databases/
    └── …

On a large data warehouse the metadata export can easily reach hundreds of megabytes or several gigabytes. Committing it:

• bloats the repository and slows every clone and fetch,
• produces noisy diffs on unrelated PRs whenever someone resyncs,
• can make the repo effectively unusable for CI and for new contributors.

Each developer (or a CI job) fetches metadata on demand from their own Metabase instance instead.

Each developer downloads table_metadata.json from their Metabase instance and drops it into .metadata/. Then run the extractor:

mkdir -p .metadata
# Drop table_metadata.json from Metabase into .metadata/

rm -rf .metadata/databases
bunx @metabase/database-metadata extract-table-metadata .metadata/table_metadata.json .metadata/databases

After this, tools and agents should read the YAML tree under .metadata/databases/ — not table_metadata.json, which exists only as input to the extractor.

Releases are published automatically by the Release to NPM GitHub Actions workflow on every push to main. The workflow compares the version in package.json against the version published on npm and publishes (with the latest dist-tag) if they differ.

To cut a release, bump version in package.json and merge to main.

The workflow requires an NPM_RELEASE_TOKEN secret with publish access to the @metabase npm org.

bun install
bun bin/cli.ts extract-table-metadata examples/v1/table_metadata.json /tmp/.metadata/databases

• bun run build — compile TypeScript to dist/ and bundle the spec.
• bun run type-check — tsc --noEmit.
• bun run lint-eslint — ESLint with no warnings allowed.
• bun run lint-format — oxfmt format check.
• bun run test — bun test suite.

The Lint, Test, and Validate GitHub workflows run on every push and pull request. Validate regenerates the bundled examples and fails if they drift from what's checked in.