I Replaced Metabase With Claude. Here's What That Looks Like in Practice.

I get the "is AI going to replace Metabase" question almost every week. The honest answer is "for a lot of teams, yes — but not in the way most people are pitching it."

I want to walk through the actual migration I did, from a Metabase install I'd run for two years to a workflow that's mostly Claude + QueryBear + a single dashboard. What I kept, what I dropped, what I learned.

The Metabase install I had

For two years I ran Metabase on a $20/month box with a connection to my Postgres production replica. About 15 saved questions. Three dashboards: one for daily metrics, one for billing, one for support load.

It worked. Nobody complained. I also opened it once a week, reluctantly, because the alternative was writing the queries from scratch.

The honest numbers:

• ~12 questions accounted for 95% of all dashboard hits.
• Two of the three dashboards drifted out of date within six months of being built.
• About once a quarter, a schema migration broke a saved question and nobody noticed for weeks.
• Total cost including Metabase Cloud trial conversions and the box itself: $48/month. Not crazy. Also not free.

The migration

I didn't kill Metabase in a single Friday. I killed it gradually over about six weeks. The order ended up being:

Week 1. Connected QueryBear to the same Postgres replica. Re-asked Claude my five most-used Metabase questions, saved each one in QueryBear.

Week 2. Added the daily-metrics dashboard equivalent. This is where it got interesting. Metabase had a 6-card dashboard. QueryBear didn't have a dashboard primitive (still doesn't, deliberately). What it has is "tools" — saved queries with parameters that you can pin to your home screen. I made each metric a tool. Functionally identical experience for me, just no auto-refreshing chart with a single header.

Week 3. Discovered I wasn't missing the charts. I was looking at table output and reading the numbers. The chart was decoration.

Week 4. Wired QueryBear's MCP server into Claude Code. Now when I'm in my terminal and want to know "did anyone sign up overnight," I don't open a tab. I ask Claude. It calls the saved tool. Number comes back inline. This is the change that actually shifted my workflow.

Week 5. Killed the Metabase server. Saved $48/month.

Week 6. Wrote a single proper dashboard for the metrics I want emailed to me weekly. Used a small Python script + a cron + an SMTP send. ~30 lines of code. This is the shape of "real dashboards" that AI doesn't replace.

Where AI clearly wins

Zero setup for new questions. Metabase needs me to build a question. Claude+QueryBear answers from a question. The setup cost on the second one rounds to zero.

No maintenance. A schema rename in Metabase is a manual fix on every saved question that touches the renamed column. In QueryBear, the AI re-resolves names against the current schema. Not magic — sometimes I need to refresh the schema cache after a migration — but the cost goes from "audit every question" to "click a button."

Conversational follow-ups. "Show me revenue by month." "Now break that down by plan." "Now only for Q1 signups." In Metabase, that's three dashboard questions or one question I edit three times. In QueryBear, it's a chat. Each follow-up takes seconds, not minutes.

Costs that don't scale per-seat. Metabase Cloud charges per user past a small included pool. QueryBear is flat. For a 10-person team that's the difference between $80/mo and $800/mo. That math gets worse the bigger the team gets.

Where Metabase still wins

Honest about this, because the marketing post would skip it.

Pixel-perfect dashboards for execs. If your CFO needs the revenue chart to look the same every Monday morning, AI isn't the answer. A static, scheduled, formatted dashboard is the answer. Metabase or a Looker Studio thing or a homemade Python+matplotlib pipeline does this better.

Embedded analytics in your product. If you're embedding charts in a customer-facing dashboard, Metabase has a real story for that. AI tooling does not.

Highly visual work. Funnel charts. Cohort heatmaps. Geographical maps. AI can produce SQL for the data, but the rendering layer in Metabase is genuinely good and there's no equivalent in chat-style tools.

Recurring exports. If you need a CSV emailed to the finance team every Monday at 8am, that's a job for a thing that can be scheduled. QueryBear's tools can be re-run, but it's not a scheduler.

The 80/20 split that fell out of this

After six weeks, what I actually do looks like this:

• 80% of my "data needs" go through QueryBear + Claude. Ad-hoc questions, debugging support tickets, daily checks, looking up specific accounts.
• 15% goes through saved tools that I or a teammate run with parameters. Same surface, just faster.
• 5% goes through a real scheduled dashboard or report. The exec-facing stuff. The recurring exports.

The 80% used to all be in Metabase. It didn't need to be. It was there because I had no better place to put it.

What I'd tell someone considering the same move

Don't try to replace Metabase by building dashboard equivalents in an AI tool. That's the wrong lens. You're not migrating dashboards. You're recognizing that most of your dashboard usage was actually ad-hoc question usage in disguise, and that the ad-hoc question usage is what AI is genuinely good at.

The tiny remaining set of "yes, this is a real dashboard that runs on a schedule" — keep using something for it. Could be a stripped-down Metabase. Could be a 30-line cron + SMTP script. Could be Looker Studio if you live in Google's world.

For everything else: chat to your database, save the question if it's useful again, and stop treating dashboards as the only legitimate way to look at numbers.

The best replacement for most BI is "ask the database directly, with a gateway in front of it." That gateway is QueryBear. The chat is whichever LLM you trust. The dashboard is the small remaining surface that actually deserves to be a dashboard.