Metabase

Meetabase Berlin 🇩🇪: with only a week's notice, we filled the room (packed staircase included). Our CEO, Sameer Al-Sakran, chatted with Francesco Mucio about what it takes to build an open source company, AI in analytics, and the future of BI. Thanks to everyone who came and to dltHub for co-hosting. We'll do it again soon ✌️

How do you know you’re living in June 2026 and not June 1999? A. Knicks vs Spurs in the 🏀 NBA finals B. World Cup is about to kick off ⚽️ in the US C. Natural-language data querying 💡 is a daily (hourly?) activity We know AI is now integral to data work, so we include it in every version of Metabase. And we’re building infra and shipping features that make AI-generated insights accurate and reliable. Post about this ↓

The future of BI is not better dashboards. It's your AI agent querying your data directly. I connected Claude Code to our Metabase instance last week. One prompt. 8 seconds. Finished card. Тo SQL, no editor. Then I pushed it further: → 500+ saved questions audited in an afternoon (duplicates, stale queries, usage ranking, the stuff that takes a week of analyst time) → Weekly KPI watchdog running on a cron job, caught a 22% churn spike at 9am Monday, three days before anyone would've noticed → CEO asks "are we getting more efficient at acquiring customers?", Claude pulled 6 data sources, computed CAC payback trend, answered in 90 seconds with citations Honest caveats: it hallucinates column names. Long queries time out. Open-source Metabase loses some Pro features. This is a fast junior, not a closed loop. But that's not the point. The point is: you don't need more analysts running queries. You need someone architecting the layer that lets your whole team ask questions directly. Clean schema. Real documentation. Agent-readable data. That's the work that matters now. Recorded the full demo: 3 real scenarios, live, no faked footage: https://lnkd.in/dyibG5by Drop a comment: what's the first question you'd ask an AI about your data?

Here are the different ways to use AI with Metabase: 🔹 Metabot: for the people who live in Metabase. Ask questions, build queries, explain a chart, draft a Document. Plus admin controls over who can use what, and how much. 🔹 MCP server: for the people who live in Claude or Codex. Combine data from Metabase with data from other tools, ask ad-hoc, ephemeral questions, and have it build the dashboard too. 🔹 File-based dev: for the builders. A coding agent reads your schema, writes dashboards as YAML, you verify, then sync to prod. Version it like real code. AI exploration that runs on your permissions and your semantic definitions.

Metabase’s Sameer Al-Sakran shared a hot take on why AI agents are poised to eclipse classical BI tools during yesterday’s Data Berlin x dltHub event: 📊🥊🤖 🚀 1. Good is good enough Humans build imperfect, biased dashboards all day long. AI agents just need to be slightly more accurate and consistent than our mediocre human baseline. ⚡ 2. Decision velocity > absolute quality We already trade perfect prose for speed when using AI in writing. Analytics is next. In most scenarios, people will happily trade a higher error rate for near-instantaneous decision making. 😢 3. It will get worse before it gets better Getting started with Claude, ChatGPT, and co. is easy and exciting. But learning how to use AI Agents effectively for complex data analysis takes serious time and training. The real paradigm shift will hit once teams cross to the other end of that learning curve. 👉 High-stakes decisions where a human MUST hand-roll the numbers aren't going away entirely, but they will become the exception. Personally, I've built roughly 3x as many Warehouse-MCP servers as dashboards YTD (props to nao Labs (YC X25) for doing the heavy lifting to fix the context layer for Analytics agents, for instance). I'm genuinely interested how others are navigating this shift? What tools are you using to ground your agents in rock-solid numbers? 🤔

Our LSP took three minutes to boot. Three. Minutes. Turns out a 500K-line Clojure codebase will find every soft spot in your dev tooling. So, Oleksandr Yakushev (longtime Clojure performance specialist) went spelunking through flamegraphs and came back with Clojure-LSP's startup time cut in half and memory allocation down by two-thirds. The fixes are already merged upstream. If you use clojure-lsp on a big codebase, your next update just got faster, and you didn't have to do anything. And he also wrote up exactly how, profilers and all ↓

"Why did signups dip last week?" You can ask that directly in Codex, and get an answer grounded in the metrics your team trusts. We are now a verified plugin in ChatGPT and Codex, as part of OpenAI's new role-specific plugins. The context you've built in Metabase is available where you're already working. Learn more ↓