Metabase AI: Capabilities, Setup, and Getting Started

Yes, Metabase has AI. Metabot is a built-in assistant that writes SQL, generates charts, and answers questions in plain English without training required. It arrived in January 2026 (v58) and has evolved through three releases to support multiple LLM providers. You can run it self-hosted or cloud, and it respects your existing data permissions.

What Is Metabase AI?

Metabase added AI capabilities to help teams extract insights without writing SQL or building dashboards manually. The main feature is Metabot, launched as GA in January 2026. It's built directly into Metabase and works alongside your existing dashboards and reports. You don't need to train a model or prepare special datasets. Metabot reads your semantic layer (the business glossary and verified data model) and uses it to understand what your data means.

Metabase AI isn't a separate tool you buy. It's integrated into the application. You enable it in the admin panel, connect an LLM provider (Claude, GPT, Bedrock, Azure, Mistral, or others), and users can start asking questions immediately. The AI respects all your existing row-level and column-level permissions. Metabase open-sourced its AI capabilities in April 2026 (v60), and the latest release is v63 (July 2026).

Three ways to use Metabase AI exist:

1. Metabot (in-app chat for users)
2. MCP Server (integrate with Claude, ChatGPT, or Cursor)
3. Agent-driven dashboards (AI commits YAML dashboard files to git)

Most teams start with Metabot because it requires no setup beyond connecting your LLM provider.

Metabot: Your AI Assistant for Data

Metabot is Metabase's conversational AI. Users ask it questions in plain English, and it returns SQL queries, charts, or summaries. It's designed to handle ad-hoc exploration and self-serve analytics without asking an analyst for every question.

What Metabot can do:

• Answer natural-language questions about your data
• Generate SQL queries from descriptions
• Create and configure charts (bar, line, map, scatter, etc.)
• Summarize results and explain findings
• Suggest follow-up questions based on context

Metabot is context-aware. It knows the structure of your databases, the tables available to each user, and the semantic layer definitions you've created. A semantic layer is your business dictionary: if you've added terms like "ARR" (annual recurring revenue) or defined "active users" as a metric, Metabot uses those definitions to construct accurate queries. Without a strong semantic layer, Metabot generates SQL that's technically correct but often misses what you actually want.

One critical point: Metabot doesn't have access to data you haven't explicitly exposed in Metabase. If a table or column isn't visible to a user, Metabot won't query it, even if asked. This is intentional. It means your permissions model is enforced end-to-end. An analyst with access only to US regions can't accidentally ask Metabot about European data.

Connecting Metabase AI to Your LLM Provider

Metabase doesn't host its own LLM. You bring your own provider and API key. This is how it works step by step:

1. In Metabase admin panel, go to Settings > AI settings.
2. Choose your LLM provider. Supported options are:
   • Anthropic Claude (including Claude 3.5 Sonnet)
   • OpenAI GPT-4 and GPT-4 Turbo
   • AWS Bedrock
   • Azure OpenAI
   • Mistral AI
   • OpenRouter
   • Together AI (Z.AI)
3. Paste in your API key from your chosen provider.
4. Configure token limits and usage auditing if needed (optional).
5. Test the connection by asking Metabot a question.
6. Enable Metabot for specific users or groups, or enable it globally.

There's no lockdown to a single provider. You can switch providers at any time, experiment with different models, or run Metabot for some users and not others. Metabase tracks usage, so you can see exactly how many tokens each conversation consumes.

Costs depend on your choice. If you use Claude via Anthropic's API, you pay Claude's pricing: approximately $3 per million input tokens and $15 per million output tokens (as of v63). If you use GPT-4, you pay OpenAI's rates. Bedrock and Azure have their own pricing structures. The flexibility is intentional: you pick the model that fits your cost and performance requirements.

One note on compliance: Metabot sends your database schema and some sample data to your LLM to generate correct queries. This is normal LLM behavior, but it's important to understand if your data has compliance constraints (HIPAA, GDPR, etc.). Most enterprise LLM providers offer data residency options to address this concern.

How AI Uses Your Data Model

The quality of Metabot depends entirely on the quality of your semantic layer. The semantic layer is the business dictionary that sits on top of raw database tables.

If you haven't built a semantic layer, Metabot works but performs like a junior analyst: it can generate SQL, but it might misinterpret what users want. For example, if a table has a column named usr_rtn instead of user_retention_percent, Metabot has to guess. It might generate technically correct queries that answer the wrong question.

If you have a well-built semantic layer, Metabot becomes a power tool. You define:

• Table and column aliases and descriptions
• Metrics (revenue, churn rate, conversion rate)
• Dimensions (by country, by region, by product)
• Relationships between tables and join logic
• Business rules and formula definitions

Metabase then passes all of this context to the LLM before each query. Metabot can ask clarifying questions ("Do you mean gross margin or net margin?") and propose follow-ups ("Would you like to see this broken down by product line?"). It becomes a conversation, not just a SQL generator.

This is why the Metabase community recommends starting with a semantic layer before rolling out AI widely. It's not required, but it's the difference between "AI that generates answers" and "AI that generates correct answers." For more on how to plan a Metabase deployment, see our guide to Metabase pricing and plans.

Three Ways to Use AI with Metabase

Metabase gives you three distinct paths depending on your workflow:

Metabot (In-App Chat)

Users click the Metabot icon in Metabase and start asking questions in plain English. Answers appear as runnable queries and auto-rendered charts. This is the most accessible route for non-technical users and business analysts. No setup required beyond enabling AI in admin settings. Most teams use this as their primary mode.

MCP Server (Claude, ChatGPT, Cursor)

Model Context Protocol allows you to connect Metabase to Claude, ChatGPT, or Cursor as a plugin. Users never open Metabase; they ask Claude about their data directly. Cursor users (an IDE) can ask about data right in their editor while they code. This requires some configuration: your Metabase instance needs to be accessible from the internet, and you need to register the MCP server with your LLM platform. This approach integrates AI into workflows that already use Claude or ChatGPT.

Agent-Driven Dashboards (Dashboards as Code)

For developers, you can let AI agents commit dashboard YAML files to a git repository. This is the most programmatic approach: AI helps you build dashboard infrastructure, not just query data. You'd use this if you're managing dozens of dashboards and want AI to help automate dashboard creation and versioning. This is the least common starting path and requires the most technical setup.

Most teams start with Metabot (option 1) and expand from there.

Permissions and Security: AI Respects Your Rules

A core guarantee of Metabase AI: Metabot cannot see data a user isn't supposed to see. This is critical for multi-tenant systems or companies with sensitive data in the same database.

Metabase's permission model includes:

• Row-level permissions (database-wide, table-wide, or row-based on data attributes)
• Column-level permissions (users can see tables but not specific columns)
• Group-based access (permission bundles for teams)
• Admin roles with view-only, edit, or full control tiers

When you enable AI, all of these permissions flow through. If user Alice can only see North American customers, Metabot answers her questions using only North American data. If Bob can't see the salary table, Metabot won't query it for him, even if he asks. This is not a feature that must be configured; it's automatic.

Usage auditing is built in. You can see:

• Which users are using Metabot and when
• How many questions they've asked
• How many tokens each query consumed
• The actual prompts and responses (in admin settings)

You can also set per-user or per-group token limits. For example: "Analysts can use 1 million tokens per month, but executives can use 5 million." This is a cost control mechanism and a way to prevent accidental overuse.

Planning Your Infrastructure

Self-hosted Metabase supports AI. Cloud-hosted Metabase (Metabase Cloud) supports AI. Both work the same way: you connect your LLM provider in settings, and Metabase handles the integration.

For self-hosted, you need:

• Outbound HTTPS access (to reach the LLM API: Anthropic, OpenAI, etc.)
• A secure way to store and rotate API keys (environment variables or a secrets manager)
• Network isolation if your data is sensitive (your LLM provider shouldn't see internal IP addresses unless you want it to)

Metabase Cloud is simpler: Metabase manages all infrastructure, and you only configure the LLM connection. They handle scaling, backups, updates, and networking. You pay a subscription fee ($50-$500/month depending on usage and team size) plus your LLM provider costs.

Costs break down into two categories:

1. Metabase infrastructure (self-hosted = your server/database costs, Cloud = subscription fee)
2. LLM provider costs (Claude, GPT-4, Bedrock, etc., billed by token consumption)

For a 20-person team asking perhaps 50 questions per day via Metabot, expect $20-$50/month in LLM tokens. This is based on roughly 10k tokens per question (input + output), 50 questions/day, approximately 1.5 million tokens per month at current Anthropic pricing.

See Metabase hosting options to compare self-hosted and managed deployment paths.

Frequently Asked Questions

How do you enable AI in Metabase? Go to admin Settings > AI, choose your LLM provider (Claude, GPT-4, Bedrock, Azure, Mistral, OpenRouter, or Z.AI), paste your API key, and save. Test the connection. Users can then access Metabot immediately (if you've granted them permission). It takes about 5 minutes.

Can you use Metabase AI with self-hosted deployments? Yes. Whether you run Metabase on your own servers or use Metabase Cloud, you can enable AI by connecting an LLM provider in settings. Self-hosted and Cloud work identically for AI features.

What LLM providers does Metabase support? Metabase supports Anthropic Claude, OpenAI GPT-4, AWS Bedrock, Azure OpenAI, Mistral AI, OpenRouter, and Together AI (Z.AI). You can also bring any provider that offers an OpenAI-compatible API.

Does AI respect my data permissions? Yes. Metabot only queries data that the logged-in user is allowed to see. Permissions are enforced end-to-end. Row-level, column-level, and group-based permissions all apply automatically to Metabot queries.

Can I use Metabase AI without a semantic layer? Yes, but it works significantly better with one. Without a semantic layer, Metabot is like a junior analyst making educated guesses. With one, it's a senior analyst who understands your business definitions.

How much does Metabase AI cost? Metabase doesn't charge a separate fee for AI. You pay for your LLM provider based on tokens consumed. Typical cost for a small team: $20-$100/month. Metabase Cloud adds a subscription fee ($50-$500/month depending on scale).

Can I disable AI for certain users or groups? Yes. In admin settings, you can enable AI globally or per-group. You can also set token budgets per user or group to limit spending and prevent overuse.

The Bottom Line

Metabase AI (Metabot) is production-ready and has been since January 2026. It works best with a strong semantic layer; without one, accuracy suffers. You control costs by choosing your LLM provider and setting token budgets. Permissions flow through automatically.

If your team asks lots of ad-hoc questions, Metabot saves analyst time and enables self-serve. If your data model is weak or your semantic layer is missing, Metabot will generate incorrect answers. Start with a solid semantic layer, then roll out Metabot to users.

Ready to run Metabase with AI? Opsily's managed Metabase hosting handles the infrastructure, so you focus on your data and data model.