How to Set Up Metabase Metabot AI

Metabase Metabot AI lets you ask plain-English questions about your data and get charts, SQL queries, and summaries without writing a line of code. Setup takes under 30 minutes on Metabase Cloud or self-hosted. You need an Anthropic API key for the self-hosted path. Once connected, Metabot can generate charts, fix SQL errors, and analyze dashboards on demand.

What Metabot Actually Is

Metabot is Metabase's built-in AI assistant, powered by Anthropic models. Unlike general-purpose AI tools bolted onto analytics platforms, Metabot has direct access to your database schema, your existing questions, and your business glossary. When you type "Show me monthly revenue by region since January," it does not just write SQL. It first checks whether you already have an answer to that question saved in your instance and points you there instead of creating a duplicate.

Metabase is available free as an open-source tool you host yourself and as Metabase Cloud starting at $500 per month for the Pro tier. Metabot is available on both paths, but the access method differs. Self-hosted users bring their own Anthropic API key. Cloud users can purchase Metabot from the Metabase Store or also bring their own key.

According to Metabase's own benchmarking, teams that add table and column descriptions before using Metabot see noticeably fewer irrelevant or incorrect responses compared to teams that skip that step. That documentation work is the highest-leverage thing you can do before asking your first question.

Prerequisites

Before enabling Metabot, confirm you have the following:

• A running Metabase instance (version 50 or later) with at least one database connected
• Admin access to your Metabase instance
• For self-hosted: an Anthropic API key, obtained at console.anthropic.com
• For Metabase Cloud: access to store.metabase.com linked to your instance

You do not need a data engineering background. If you can connect a Postgres database or a Google Sheet to Metabase, you can set up Metabot.

Step 1: Connect a Database

Metabot is only useful when it has data to work with. If you have not already connected a data source, do that first.

1. Go to Admin > Databases > Add a database
2. Choose your database type. Metabase supports PostgreSQL, MySQL, BigQuery, Redshift, MongoDB, DuckDB, and others. Google Sheets can be connected via the official partner connector.
3. Enter credentials and click Save

Metabase will sync the database schema automatically. For large databases this can take several minutes. Wait until the sync shows "Complete" in your admin panel before moving on. If Metabot does not have an up-to-date schema, its answers will be less reliable.

Step 2: Enable Metabot

For Metabase Cloud:

1. Go to store.metabase.com and log in with your Metabase Store account
2. Find your instance in the Instances tab
3. Click Add Metabot AI and select a plan based on your expected request volume
4. Accept the terms and complete the purchase
5. Return to your Metabase instance and navigate to Admin > AI

For self-hosted Metabase:

1. Go to Admin > AI > Connection settings
2. Select Anthropic from the provider dropdown (it is currently the only supported provider)
3. Paste your Anthropic API key
4. Click Connect
5. Select a model from the dropdown. Available models are fetched from your API key in real time.

Both paths land you in the same Admin > AI settings panel. Toggle Enable Metabot on to make it visible to all users in your instance.

Step 3: Configure Permissions and Scope

Metabot inherits the data permissions of whoever is chatting with it. A viewer can only ask about data they already have access to. An admin asking a question gets answers drawn from tables restricted to admins. No separate permission layer is required, and no special Metabot roles exist.

Two optional settings are worth enabling immediately:

Verified content only: restricts Metabot to models, metrics, and questions that an admin has explicitly marked as verified. This reduces cases where Metabot invents column names or joins tables incorrectly. Available on Pro and Enterprise plans.

Collection limits: scopes Metabot's exploration to specific collections instead of the entire database. If your ops team works from a separate collection than your finance team, scoping prevents cross-contamination of answers and keeps responses focused.

Step 4: Prepare Your Data

This step is optional to skip but not optional if you want good results. Metabot performs significantly better when it knows that a column named ord_rev_usd means "order revenue in US dollars" and not something else.

Add table and column descriptions:

1. Go to Admin > Data Model
2. Select a table
3. Add a description in plain English. Example: "Contains one row per completed customer order, with revenue, region, product category, and order date."
4. Add field descriptions for any column with a non-obvious name. Example: "net_rev: revenue after refunds, not including shipping or tax"

Build a Business Glossary:

1. Go to Admin > AI > Glossary
2. Add terms your team uses internally. Example: "MRR: Monthly Recurring Revenue, sum of all active subscription values divided by 12"
3. Add acronyms and metric definitions specific to your organization

This glossary is fed directly to Metabot when it processes your queries. Based on Metabase's documentation, a well-structured glossary improves answer relevance more than any other single configuration step, because it closes the gap between how your team talks about data and how the database actually stores it.

Step 5: Ask Your First Question

Open the Metabot sidebar with Cmd+E on Mac or Ctrl+E on Windows. You can also click the Metabot icon in the top-right corner of any page.

Start with a concrete question tied to something you actually need:

• "Show me total orders per month for the last six months"
• "Which customers have not placed an order in 90 days?"
• "What is our average order value by product category this quarter?"

Metabot checks your saved questions and dashboards for a match before building anything new. If it finds one, it links you there. If not, it uses the query builder to construct a chart.

For SQL-heavy work, open the native query editor and press Cmd+Shift+I on Mac or Ctrl+Shift+I on Windows to trigger inline SQL suggestions. Type your intent as a comment and Metabot will write the SQL below it.

If you want to skip the database connection and glossary overhead entirely, VSLZ AI handles natural language queries directly from a file upload, with no schema configuration or column descriptions required.

Step 6: Save and Share Results

Any chart Metabot generates can be saved directly to a collection or pinned to a dashboard. After reviewing the result, click Save, choose a destination collection, and the chart becomes part of your permanent analytics setup. Other team members with collection access can view, edit, and build on top of it.

From any saved chart, you can configure an alert in Metabase to notify your team by email or Slack when a metric crosses a threshold. This turns Metabot's one-off answers into ongoing monitoring with no additional tooling.

What Metabot Cannot Do

Knowing the limitations saves troubleshooting time:

• Metabot cannot use custom expressions or multi-level aggregations
• It cannot modify chart formatting including colors, axis labels, and goal lines
• It cannot generate queries that rely on SQL parameters
• It does not perform well in languages other than English
• It cannot discover segments or saved documents on its own
• It cannot edit alert configurations or delete saved questions

For analysis involving complex nested aggregations or multi-step data transformations, a data analyst reviewing and adjusting the generated SQL is still the right call.

Practical Summary

Setting up Metabase Metabot AI involves four substantive steps: connect a database, enable Metabot with an Anthropic API key, restrict to verified content and relevant collections, and write descriptions for your key tables and columns. The quality of Metabot's answers scales directly with the quality of your data documentation. Teams that invest 30 to 60 minutes in the Admin > Data Model and glossary setup before inviting users to Metabot report far fewer follow-up questions and fewer wrong answers compared to teams that skip it.