How to Set Up AI Features in Metabase 60

Metabase 60, released in April 2026, is the first version of Metabase to include AI features across all plans, including the free open-source build. Previously, AI assistance was locked to Pro and Enterprise tiers. The update opens natural language querying, SQL generation, chart summaries, an official MCP server, and Metabot in Slack to anyone running a self-hosted instance or using Metabase Cloud.

This guide covers the complete setup for the Bring Your Own Key (BYOK) flow, which works on all plans, and walks through each feature so you know what to enable for your team.

What Metabase 60 Added

The April 2026 release shipped five distinct AI capabilities:

• Metabot: A natural language assistant that creates charts, generates SQL, fixes query errors, and summarizes existing visualizations on demand.
• MCP server: An official Model Context Protocol endpoint at /api/mcp that lets Claude, Cursor, VS Code, and ChatGPT query your Metabase data directly.
• Metabot in Slack: Ask data questions, upload CSVs, and set up alerts inside a Slack channel without opening Metabase.
• Agent API: A programmable interface for building custom AI workflows on top of your Metabase data.
• AI SQL generation: Inline suggestions and one-click error fixes inside the SQL editor.

The release also introduced split panel charts (multi-series charts split into panels with independent y-axes), a metrics explorer for side-by-side comparisons, and semantic search that understands meaning beyond exact keywords.

Step 1: Navigate to AI Settings

Log into your Metabase instance as an admin. Go to Admin settings (the gear icon, top right) and click AI in the left sidebar.

You will see two connection options:

• Metabase AI service: Metabase picks a cost-effective model and bills through your Metabase account. Available to Cloud users at no extra charge and is the faster path for Cloud teams.
• Bring your own API key: Connect your Anthropic key directly and pay per token. This option works on all plans, including the open-source self-hosted build.

For self-hosted instances on the open-source plan, the BYOK route is the only available path. Cloud users can use either option.

Step 2: Connect Your Anthropic API Key

Select Your own API key, then complete these steps:

1. Click the link provided to open your Anthropic Console. Generate a new API key or copy an existing one.
2. Paste the key into the field in Metabase and click Connect.
3. Select a model from the dropdown. The list reflects models available on your Anthropic account tier.

A green "Connected to Anthropic" badge confirms the key is live. Metabase currently supports only Anthropic models through the BYOK flow. Support for additional providers has not been announced as of April 2026.

For cost reference: at Claude Haiku pricing, a team running 50 to 100 Metabot queries per day typically spends under $5 per month on token costs. Teams using Claude Sonnet for more complex SQL generation should budget $20 to $50 per month depending on query complexity and volume.

Step 3: Configure Metabot

After connecting, open the Metabot tab to configure how the assistant behaves:

Toggle Metabot on. There are two independent switches: one for internal Metabase users and one for embedded analytics contexts. Enable only what applies to your deployment.

Verified content filter. When enabled, Metabot only searches models and metrics that have been verified by your data team. This prevents it from surfacing unfinished or experimental tables in answers. For production environments, this setting should be on.

Collection limits. Specify which collections Metabot can access when responding to questions. Scoping this to the two or three collections that matter for a given team keeps answers focused and avoids Metabot pulling from irrelevant datasets.

Prompt suggestions. Metabase auto-generates suggested questions based on your most-used models and metrics. These appear as clickable prompts in the Metabot panel. Useful for non-technical users who are unsure how to phrase a question.

Once enabled, users access Metabot from the question editor or dashboard view by clicking the Metabot icon. They can type questions in plain English, paste in error messages from failed SQL queries, or click the auto-summarize button on any existing chart.

Step 4: Enable the MCP Server

The MCP server lets external tools connect to your Metabase and query data without requiring users to log in through the browser. All access is scoped to the connecting user's existing Metabase permissions.

To enable it:

1. In Admin > AI, find the MCP Server section and toggle it on.
2. Select which MCP clients are permitted to connect: Claude, Cursor, VS Code, ChatGPT, or all of them.
3. For self-hosted instances, add your domain to the allowed domains list. Wildcard subdomain entries are supported.

Your MCP endpoint is [your-metabase-url]/api/mcp. Point any MCP-compatible client at that URL. The client authenticates using the user's Metabase credentials, so a user who cannot see a particular collection in Metabase cannot access it through the MCP connection either.

For teams already using Claude or Cursor as a daily working tool, the MCP server is the highest-value feature in this release. Analysts can query live Metabase data without switching applications.

Step 5: Set Up Metabot in Slack

The Slack integration lets team members ask data questions and get answers inside Slack without opening Metabase. It also supports CSV uploads: send a file in Slack and follow up with questions about it.

Setup steps:

1. In Admin settings, go to the Slack section (separate from the AI tab).
2. Click Connect Slack and authorize the Metabase Slack app for your workspace.
3. Invite Metabot to channels where you want it active, using /invite @Metabase.

Once active, users can ask questions by mentioning Metabot directly in a channel. CSV uploads trigger an automatic parsing confirmation, after which follow-up questions work in the same thread.

One practical limit: Metabot in Slack responds based on the collection limits and verified content settings you configured in Step 3. If a dataset is not in an allowed collection, Slack queries will not surface it.

Common Setup Issues

"Connected" badge does not appear: Double-check that the Anthropic API key has not been restricted to specific IP ranges in your Anthropic Console. Metabase cloud servers and self-hosted instances both need unrestricted key access.

Metabot returns empty results: This usually means the collection limits are too narrow, or the tables Metabot needs have not been verified. Check both settings in the Metabot configuration tab.

MCP client cannot reach the endpoint: Confirm your Metabase domain is on the allowed domains list in the MCP settings, and that your self-hosted instance is accessible from the machine running the MCP client. Localhost-only deployments require the MCP client to run on the same machine.

Slack integration is active but Metabot does not respond: The Metabase Slack app requires the channels:read and chat:write scopes. If the app was installed before Metabase 60, you may need to reinstall it to pick up the new permission requirements.

What to Turn Off

Not every feature should be on for every deployment:

• Turn the MCP server off if your team does not use Claude, Cursor, or other MCP clients. There is no benefit to leaving it on if no clients are configured.
• Keep Metabot in Slack off in channels with sensitive data unless you have confirmed that collection limits exclude those datasets.
• The Agent API is designed for developers building custom integrations. Leave it off unless you have a specific use case requiring programmatic AI access to Metabase.

The master "Disable all AI features" toggle in Admin > AI shuts everything off instantly without disconnecting your Anthropic key. Use it if you need to pause AI access during an incident or review period.

Practical Next Step

The fastest way to validate the setup is to open a question that already has data, click the Metabot icon, and type "summarize this." If you get a plain-English interpretation of the chart, the connection is working. From there, test natural language querying with a question about one of your verified metrics.

For teams that want to avoid managing API keys, model selection, and token budgets altogether, VSLZ AI lets you upload a data file and ask questions in plain English without any configuration.