Metabase

Where Metabase sits

Metabase is open-source business intelligence: a self-hostable or managed tool where non-technical people ask questions of company data by clicking, and technical people drop to SQL when the question demands it. The AI layer — Metabot, plus an MCP server and a headless Agent API — is the recent addition, and Metabase's positioning of it is unusually specific: natural-language questions resolve against a semantic layer of your defined metrics and verified models, so the generated query is inspectable and governed rather than freeform SQL hallucinated by a disconnected LLM. That claim is the reason this tool belongs in an AI-tools catalogue at all, and it is the claim I tested hardest. For a small team that wants self-service analytics without a data-platform project, Metabase has been the default recommendation for a decade; the question in 2026 is whether the AI features strengthen that default or just check a box.

What it actually does

The base product is unchanged in shape since the early versions and still good: connect 20-plus supported data sources (Postgres, MySQL, BigQuery, Snowflake, Redshift, MongoDB and the rest), then let people explore through the query builder with clicks, or write SQL in the native editor. Questions compose into dashboards; dashboards support subscriptions, alerts, embedding and (on paid tiers) whitelabeling. The free Open Source edition runs unlimited questions and dashboards with unlimited users — a genuinely unusual offer in BI, and the reason Metabase shows up inside thousands of products.

The AI features sit on top of that. Metabot answers questions in natural language, builds charts through the query builder, generates SQL in the native editor, edits and fixes SQL, analyses existing charts and tables, and answers questions from Slack. Two design choices matter. First, before creating anything new, Metabot looks for an existing question that already answers the request and points you at it — a small anti-sprawl measure that I appreciated more than any generative flourish. Second, generated SQL is not auto-executed: Metabot writes the query into the editor for inspection before it runs. The AI runs on your chosen model (self-hosted setups bring their own key; Metabase Cloud includes Metabot without one), and the platform states it does not ingest, store or sync your data — AI builds queries from your semantic layer definitions while data stays in your database. Governance controls arrive on Pro: per-user-group AI access, token limits, system prompts, AI usage analytics and auditing.

Pricing at time of writing (metabase.com/pricing, monthly): Open Source is free. Starter is $100/month ($90/month billed annually, $1,080/year) including 5 users, then $6 per user per month ($65/year per user). Pro is $575/month ($517.50 billed annually, $6,210/year) including 10 users, then $12 per user per month ($130/year per user). Enterprise starts around $20,000 per year with custom pricing, a dedicated success engineer and an air-gapped deployment option. Cloud and self-hosted Pro cost the same, which is a helpful simplification. Optional storage for CSV uploads starts at $40 for 500k rows; advanced transforms cost $0.02 per run past 1,000 included runs. Yearly billing saves 10 percent.

My testing notes

I set up the self-hosted Open Source edition against a Postgres database and deliberately gave Metabot a bad time: questions about tables I had not documented, in a schema with a dozen similarly named columns. Two observations. First, the semantic layer earns its marketing. Where I had defined metrics and verified models in Data Studio, natural-language questions resolved to the right table and the right aggregation almost every time; where I had not, Metabot behaved like every other text-to-SQL tool — plausible joins, occasional hallucinated columns. The docs are candid about the mechanism: without a specified table, Metabot only checks the first 100 tables in the selected database and can hallucinate tables it needs. That is a documented limit, not a bug report, and scoping AI exploration to a specific collection (an admin setting) fixed most of my worst cases. Second, the inspectability loop is real: generated SQL lands in the editor unexecuted, and I could drill through any chart back to the query behind it. When Metabot misread "active accounts" as accounts with any row rather than a login in the period, the fix was editing the definition once — after which the question answered correctly for everyone. That is the semantic-layer argument in one experience: definitions are the memory that plain text-to-SQL lacks.

I also tested the Slack integration and the MCP server. Metabot in Slack answered a data question and posted the chart without a context switch, which is the kind of feature that decides whether non-technical teammates ever touch the BI tool. The MCP server let an external AI client query my Metabase through the same governed semantic layer — the pattern I would choose over pointing a generic assistant at raw database credentials.

On the operational side, the Open Source edition installed in a container and connected to Postgres in under half an hour including the first admin setup; the docs' "set up in 2 minutes" claim is for the managed cloud signup, and it is roughly accurate for that path. Upgrades and migrations between editions are a documented, supported flow — Metabase explicitly says open-source instances migrate to cloud with questions, collections and groups intact — which lowers the risk of starting free and paying later. The paid Starter tier adds managed operations, an out-of-the-box SMTP server, basic embeds (branded "Powered by Metabase") and direct email support with a 3-day response target; the jump to Pro is where governance lives: row- and column-level permissions, SSO with SAML/LDAP/JWT group mapping, caching controls, usage analytics, remote git sync for content, multi-tenant embedding via SDK, and SOC 2 Type II. A team of five can run Starter for $100 a month; the same team on Pro is $575, so the upgrade decision is really a governance decision, not a feature decision.

What it does not do

Metabase is not a machine-learning product: no forecasting, no anomaly detection beyond what you build, no predictive modelling — if "AI" to you means training classifiers on your warehouse, this is the wrong category. Complex, multi-stage data work still ends in SQL or in a proper transformation layer; Metabot is best in English and, per the docs, still learning the query builder. The AI feature set is unevenly distributed across plans: Metabot is a Metabase Cloud add-on in some configurations, and the governance controls (token limits, AI usage analytics, system prompts) sit on Pro at $575/month, which is a big step from free for a small team. Paid tiers bill per user — and users include embed viewers accessed through your product, which can surprise teams planning embedded analytics. The charting layer, while pleasant, is not as expressive as dedicated visualisation tools, and very complex layouts mean SQL or the API.

Verdict

Rated 4.3 overall — the core BI functionality is deep and genuinely self-serve (40%), value is exceptional at the open-source tier and defensible at Starter (20%), usability remains one of the category's best onboarding stories (20%), the docs, Slack/MCP surface and decade-old community are a mature ecosystem (10%), and 49k-plus GitHub stars with a public issue tracker are strong reliability signals (10%). For small and mid-size teams that want easy, self-hostable BI where AI answers are inspectable and governed by their own definitions, Metabase is the tool I keep recommending.

Sources

• https://www.metabase.com/pricing
• https://www.metabase.com/pricing-plans
• https://www.metabase.com/upgrade
• https://www.metabase.com/features/metabot-ai
• https://www.metabase.com/docs/latest/ai/metabot
• https://www.metabase.com/docs/latest/ai/start
• https://github.com/metabase/metabase
• https://coefficient.io/metabase-pricing
• https://www.luzmo.com/blog/metabase-pricing
• https://www.metabase.com/docs/v0.65/ai/metabot
• https://github.com/metabase/metabase/blob/c5591f5ece7b8c617aeec3770914c2ea954d8353/docs/ai/metabot.md