Metabase Representation Format

Metabase Representation Format

Metabase represents user-created content as a tree of YAML files. Each file is one entity (a collection, card, dashboard, etc.). The format is portable across Metabase instances: numeric database IDs are replaced with human-readable names and entity IDs.

The format is defined by a spec bundled alongside this file as spec.md (upstream source: the @metabase/representations npm package). The same package ships a CLI (npx @metabase/representations validate-schema) that validates a tree of YAML files against the format.

Entities

The format defines 11 entity types.

Ownership and hierarchy

Critical — folder layout is decorative. Where an entity lands in Metabase is decided entirely by its fields, not by where its YAML file sits in the tree. Moving a file without updating the fields changes nothing. Updating the fields without moving the file still works correctly. Always treat the fields below as the source of truth.

The fields that actually determine placement:

• `collection_id` (entity_id of a collection) — places the entity in that collection. null or omitted → root collection.
• `parent_id` on a collection — this, and only this, sets the collection's own parent. A collection's position in the folder tree is ignored on import; without parent_id (or with parent_id: null) the collection becomes a root-level collection, no matter how deep its folder is nested. To nest one collection under another, set parent_id to the parent collection's entity_id.
• `dashboard_id` / `document_id` on a card — nests a card under a dashboard or document. Such a card must also set collection_id to match the parent's collection_id. A card never sets both.

On disk, cards nested under a dashboard or document live in a subfolder next to the parent YAML (e.g. my_dashboard/card.yaml sitting next to my_dashboard.yaml) — but again, this is purely for human navigation; the fields are what Metabase reads.

Import paths

Metabase only imports YAML from these top-level directories; anything outside is ignored:

• collections/ — all user content (cards, dashboards, documents, snippets, transforms, etc.), partitioned by namespace: main/, snippets/, transforms/.
• databases/ — only the segments/ and measures/ subdirectories under each table are imported.
• python_libraries/ (also accepted as python-libraries/).
• transforms/ — contains transform_jobs/ and transform_tags/.

serdes/meta

Every entity carries a top-level serdes/meta array that encodes its identity path. Each entry is {id, model, label?} — label is the slugified name and is present on entities keyed by NanoID. Example:

serdes/meta:
- id: NDzkGoTCdRcaRyt7GOepg
  label: my_entity_name
  model: Card

validate-schema reads serdes/meta to determine which entity type each file represents. The full rules (including nested entities and composite identity paths) are in spec.md.

Reading the spec

This skill ships with a local snapshot of the spec as spec.md alongside SKILL.md.

Beyond the per-entity shapes summarized in this SKILL, spec.md also covers: MBQL query form (stages, field references, joins, expressions, aggregations, filter/expression operators, temporal bucketing, binning), native queries and template tags (text, number, date, boolean, dimension, temporal-unit, card, snippet, table), visualization settings, click behavior, and dashboard/card parameters. Reach for spec.md whenever edits touch any of those.

Read on demand, not eagerly. Open spec.md only when you are about to read or modify content files for the entities listed above — e.g. the user asks to edit a card, add a dashcard, tweak a transform, or similar work that implies YAML edits. Do not open it at session start or for tasks unrelated to representation YAML.

If the bundled copy looks out of date with the upstream package, the skill's own README.md documents how to refresh it with extract-spec.

Validating

Validate edits with the built-in CLI:

npx @metabase/representations validate-schema --folder <path>

Pass the top-level export folder, or the git repository root. The tool walks the import paths listed above, reads serdes/meta on each file to pick the right validation rules, and exits non-zero on failure. Prefer running this over manually cross-checking field shapes. It's essentially instant, so invoke it whenever useful — after each edit, between edits, whenever the shape of a file feels uncertain. No reason to batch.

Generating entity IDs

Every entity needs a 21-character NanoID for entity_id. Generate one (or several) with the bundled CLI:

npx @metabase/representations generate-entity-id
# → LZfXLFzPPR4NNrgjlWDxn

npx @metabase/representations generate-entity-id --count 5

Generating UUIDs

Some fields in the format require v4 UUIDs rather than NanoIDs — notably lib/uuid on MBQL aggregation clauses (referenced from order-by and later stages) and the id on dashboard/card parameters. Generate them with:

npx @metabase/representations generate-uuid
# → 1d4e9fdf-49ae-4fbe-ae27-05e7c6a5cfe8

npx @metabase/representations generate-uuid --count 3

metabase-representation-format (skill)

This skill ships a local snapshot of the Metabase Representation Format specification as spec.md, next to SKILL.md. The SKILL references it directly so an agent loading this skill never has to fetch the spec on its own.

Refreshing the bundled spec

When the upstream format changes, refresh the bundled copy by running this from inside the skill folder:

npx @metabase/representations extract-spec --file ./spec.md

Commit the regenerated file alongside SKILL.md.

Files

• SKILL.md — the skill itself. Read by the agent.
• spec.md — specification (entity shapes, MBQL/native query form, folder structure, etc.). Agent reads this on demand when it needs detail beyond what SKILL.md summarizes.