Metabase Automation

Automate Metabase content, promotion, embedding, and admin refresh workflows.

Classic Metabase REST API still owns cards, dashboards, collections, permissions, and schema refresh operations. The newer Agent API is the right surface for headless semantic BI assistants and app-side AI workflows. Metabase v60 (April 2026) added an official MCP server and open-sourced AI; v61 (May 2026) added AI governance, dashboards-as-code via MCP, and per-group Metabot controls. v62 (June 2026, current line) added the official @metabase/cli, an Interactive Schema Viewer, a custom-visualization plugin SDK, an Alert Management hub, Library sub-collections, and expanded MCP capabilities — run SQL, create collections, and render interactive charts directly in the AI client. Verify /docs/latest and metabase.com/releases before citing version-specific behavior, since the release cadence is monthly.

Quick Reference

Decision Tree

Need to create or edit saved Metabase content?
  -> Use classic REST API (`card`, `dashboard`, `collection`).

Need repeatable dev -> prod promotion with reviewable diffs?
  -> Prefer Remote Sync or serialization before raw REST upserts.

Need embedded analytics for many customers or workspaces?
  -> Use embedding + tenants + embedding permissions.

Need an AI assistant to discover metrics/tables and construct queries?
  -> Use the versioned Agent API, not raw card CRUD.

Need to generate or edit questions/dashboards from an AI coding agent or terminal?
  -> Use the official Metabase MCP server (v60+, connects Claude/Cursor/VS Code).

ASCII Flow

Metabase automation request
  -> health check and authentication
  -> discover IDs: database, collection, table, fields, entities
  -> choose surface
     +-- cards, dashboards, collections -> classic REST API
     +-- dev-to-prod promotion -> Remote Sync or serialization
     +-- customer analytics -> embedding + tenants + permissions
     +-- semantic assistant -> Agent API
     +-- AI terminal / agent build -> MCP server (v60+)
  -> export existing JSON when structure is complex
  -> upsert or promote content
  -> refresh metadata if schema changed
  -> validate by running/exporting results

Quick Start

Inputs (env vars)

• METABASE_URL (e.g., https://metabase.example.com)
• Preferred: METABASEAPIKEY
• Optional: METABASE_SESSION
• Fallback: METABASEUSERNAME + METABASEPASSWORD

Sanity checks

python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py health
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py whoami
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py list-databases
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py list-collections --tree

Live API documentation

Your Metabase instance serves OpenAPI docs at /api/docs (for example https://metabase.example.com/api/docs). Use this to confirm request shapes for your exact build before scripting bulk edits.

Workflow

1. Confirm API availability (GET /api/util/health).
2. Authenticate with an API key first, then fall back to session auth only if needed.
3. Discover IDs instead of hardcoding them across environments:

- collectionid for save location - database id for datasetquery - table and field IDs if using MBQL

1. Choose the right surface:

- classic REST API for cards/dashboards/admin - Agent API for semantic, AI-driven querying - Remote Sync or serialization for promotion workflows

1. Create/update a card:

- prefer native SQL for stable automation - set display and visualization_settings explicitly

1. Create/update a dashboard, then place cards with explicit layout.
2. Refresh metadata if schema changed.
3. Validate by running/exporting results and re-opening exported JSON.

Key Concepts

• UI "Question" == API card
• Chart configuration lives on the card as display + visualization_settings
• Most visualization keys are easiest to manage by exporting an existing card JSON, then editing that payload
• Remote Sync is for Git-backed promotion, not per-request runtime automation
• Agent API is a separate, versioned surface for semantic query construction and execution; returns max 200 rows per request; paginate via continuation_token
• MCP server (v60+) connects Metabase to Claude, Cursor, VS Code, and other MCP clients with the same permissions as the authenticated user; enables dashboards-as-code from AI terminals; v62 adds SQL execution, collection creation, and interactive charts rendered in the AI client itself
• Official Metabase CLI (v62+, npm install -g @metabase/cli, distinct from the classic REST/JAR admin surfaces) builds questions, dashboards, documents, and transforms from the terminal — see references/metabase-cli.md
• Metabot (v60+, open source) integrates with Slack; v61 adds per-group access controls, token limits, and usage analytics (Pro/Enterprise only for governance features)
• Interactive/modular embedding SDK requires Pro/Enterprise, React 18/19, Node 20+, Metabase v1.52+
• Custom visualizations (v62+, Pro/Enterprise) let you build React-based chart types (Gantt, org chart, radar, calendar, heatmap) via a plugin SDK — not supported in embeds, subscriptions, or alerts

Pricing Reference (July 2026)

AI add-on: Metabase's hosted AI service bills at $3.75 per 1M tokens (1M included); most deployments instead bring their own model provider key, in which case you pay that provider directly and Metabase charges nothing extra for AI usage. Transforms include 1,000 runs/mo on Starter/Pro, then $0.01–$0.02/run.

Prices change often — verify current numbers at metabase.com/pricing before quoting costs in any proposal; do not reuse the figures above past one quarter without re-checking.

Guardrails

• Prefer Remote Sync or Metabase serialization for bulk, cross-environment promotion; use direct API for incremental upserts.
• Do not hardcode numeric IDs across environments when you can discover them or use entity IDs / synced content.
• Never commit METABASEAPIKEY, passwords, or session tokens.
• Prefer a dedicated, least-privileged automation account and collection.
• Treat internal admin notify endpoints separately from normal automation. They use server-side MBAPIKEY, not the user-facing METABASEAPIKEY.
• If a dashboard uses tabs, filters, or complex dashcard state, export first and preserve that structure instead of rebuilding from memory.

Navigation

Assets

Scripts

scripts/metabase_api.py is a dependency-free helper for auth, discovery, content CRUD, dashboard layout work, query execution, and schema refresh.

Examples:

# Print authenticated user (tries API key, then session)
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py whoami

# Discover IDs before scripting
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py list-collections --tree
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py list-databases
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py database-metadata --id 2
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py list-fields --database-id 2

# Export an existing card JSON (use as a template for visualization_settings)
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py export-card --id 123 --out card.json

# Export an existing dashboard JSON (use as a template for layout)
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py export-dashboard --id 5 --out dashboard.json

# Create/update a card from a JSON spec (see references/reports-cards.md)
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py upsert-card --spec card-spec.json

# Create/update a dashboard from a JSON spec
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py upsert-dashboard --spec dashboard-spec.json

# Add or update dashboard layout
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py add-dashcard --dashboard-id 5 --spec dashcard.json
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py update-dashcards --dashboard-id 5 --spec dashcards-layout.json

# Execute a query spec or export a saved card result
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py run-query --spec dataset-query.json
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py export-card-query --id 123 --format csv --out report.csv

# Refresh metadata after schema changes
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py sync-schema --id 2
python3 frameworks/shared-skills/skills/data-metabase/scripts/metabase_api.py rescan-values --id 2

Known Traps

• Treating Metabase as the semantic layer of record while metrics, joins, and business definitions still drift in dbt, SQLMesh, or raw SQL cards.
• Copying cards manually between environments and silently breaking references, permissions, or dashboard filter wiring because serialization or sync strategy was never defined.
• Reusing one question across dashboards with slightly different filter semantics and creating hidden KPI disagreements that look like product changes.
• Embedding tenant dashboards without a strict parameter contract, collection isolation model, and signed-embed verification path.
• Letting native SQL cards become the default authoring path for everything, then losing discoverability, lineage, and reuse across teams.
• Designing dashboard filters around visual convenience rather than stable field mappings, leading to broken linked filters after schema or model changes.

Common Anti-Patterns

• Treating dashboard screenshots or PDF exports as the decision artifact instead of versioned questions, documented metrics, and reproducible model definitions.
• Building one large "executive dashboard" that mixes raw exploratory cards, production KPIs, and operational monitoring in the same surface.
• Using personal collections as production distribution channels instead of curated shared collections with ownership and promotion rules.
• Hardcoding environment-specific table names, URLs, or card IDs into migration and automation workflows.
• Solving multi-tenancy with card duplication alone when embedding parameters, row-level source models, or separate collections would provide cleaner isolation.
• Using Metabase to compensate for unresolved source-model problems that should be fixed upstream in warehouse modeling or metric governance.

Related Skills

• [../data-analytics-engineering/SKILL.md](../data-analytics-engineering/SKILL.md) for semantic modeling, metrics, and BI-ready warehouse design
• [../data-sql-optimization/SKILL.md](../data-sql-optimization/SKILL.md) for query tuning and database performance work
• [../software-backend/SKILL.md](../software-backend/SKILL.md) for application integration and service-layer implementation
• [../agents-mcp/SKILL.md](../agents-mcp/SKILL.md) when Metabase access should be wrapped as tools rather than ad hoc scripts
• [../agents-mcp/references/mcp-for-dwh.md](../agents-mcp/references/mcp-for-dwh.md) for the LLM-via-MCP pattern catalog (Shape A managed-BI MCP applies directly to Metabase); enforcement layers (read-only role, query budgets, pseudonym boundary, audit log) apply regardless of shape

Trend Awareness Protocol

When the user asks for the current or best Metabase approach in 2026, verify the latest official docs and release notes before answering.

Trigger examples:

• "What is the best way to promote Metabase content now?"
• "Is serialization still the right choice in 2026?"
• "Should I use Agent API or classic API?"
• "What changed in Metabase 59/60/61 for embedding or dashboards?"
• "What is the current best practice for multi-tenant Metabase?"
• "How do I use the Metabase MCP server?"
• "What Metabot or AI governance features are available now?"

Fact-Checking

• Use web search/web fetch to verify current external facts, versions, pricing, deadlines, regulations, or platform behavior before final answers.
• Prefer primary sources; report source links and dates for volatile information.
• If web access is unavailable, state the limitation and mark guidance as unverified.

Learnings Loop

Before applying this skill on a non-trivial task, read learnings.consolidated.md in this directory (and learnings.md if present).

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.