Metabase shipped five GitHub releases on June 17 across v58, v59, v60, v61, and v62. That is the useful signal here. The company is not only adding AI features to a business intelligence product, it is maintaining several active release lines while trying to make governed AI analytics feel normal.

What Metabase says it is now

The official homepage frames Metabase as “analytics that answers back.” The page says teams can query in natural language, use AI-powered analytics built on metrics and permissions, bring their own LLM key, connect an AI tool through MCP, and run a self-hosted instance with a single Docker command. That is a sharper message than old BI positioning. The product pitch is not just charts for non-technical users. It is governed interaction with company data.

The GitHub repository keeps the older, broader framing: Metabase is the easy way for people in a company to ask questions and learn from data. The repository lists the expected BI surface area, no-code questions, SQL editor, dashboards, documents, alerts, subscriptions, library curation, embedding, granular permissions, dark mode, translations, and supported databases. It also now lists Metabot, an Agent API, Data Studio, Git-backed versioning, and embedded components for charts, dashboards, data browsing, and AI chat.

That combination matters. Metabase is still selling the familiar self-serve analytics promise, but the product surface is moving toward AI-native workflows. The official site says Data Studio curates the semantic layer so people and AI can trust the same metrics and definitions. The AI blog makes the same argument more bluntly: natural language querying and chart generation are becoming standard, but the hard part is the plumbing around correctness, access, limits, and auditing.

ToolVitals classifies Metabase as open core, not OSI-approved OSS. The supplied openness field says the project has an AGPL core and commercial editions. The repository license text backs that up: code outside the top-level enterprise directory is AGPL, while code inside enterprise is under the Metabase Commercial Software License unless noted otherwise. The release notes also publish separate Docker images and JAR downloads for the AGPL build and the Enterprise build. So the clean wording is open core with AGPL and commercial terms, not a generic open source label.

The metrics show a mature project that still ships hard

ToolVitals gives Metabase a hot score of 225.5, health score of 100, shipping score of 100, and overall ToolVitals score of 100. The repository has 47,723 GitHub stars. ToolVitals also records 30 GitHub releases in 90 days and 19 release events in the last 30 days, with data confidence at 100.

Those numbers say Metabase is not a sleepy incumbent. A 47,723-star BI repository could coast on reputation for a long time. This one is still producing a dense release stream. The June 17 event cluster is especially telling because it covers multiple supported versions at once: 58.17, 59.14, 60.10, 61.5, and 62.2. That looks like coordinated maintenance, not a one-off marketing release.

The release-count signal also needs a sane reading. Nineteen release events in 30 days does not prove every release contains user-visible product work. Some of those events can be patch releases, security hardening, dependency fixes, or backports across supported branches. For a BI tool connected to production data, that is still valuable. Boring release hygiene is not boring when the software sits between business users and databases.

ToolVitals does not currently show 30-day commit count or active contributor count for this payload. Those fields are null. That means the score should be read as a release and public-health signal, not a direct measurement of code churn or contributor breadth. The right conclusion is specific: Metabase is publicly active, widely starred, and shipping frequent releases. The data does not prove how many engineers touched the code this month.

Compared with related tools, Metabase sits in an interesting middle. Jitsu, another data-category tool in the payload, has 4,783 stars, a hot score of 218.0, shipping score of 95, and 40 release events in 30 days. Jitsu appears even noisier on release frequency, but Metabase has roughly ten times the GitHub stars and a perfect shipping score. LangChain, in a different developer-tools category, has 139,666 stars, hot score 239.7, shipping score 100, and 35 release events in 30 days. Metabase is not matching LangChain’s star gravity, but it is unusually active for a mature analytics product.

The current product direction is governed AI, not chatbot garnish

The Metabase 62 announcement is the clearest release source. It says v62 adds custom visualizations through a plugin SDK, interactive charts from the MCP server in AI clients, the Metabase CLI, an interactive Schema Viewer, collections in the Library, alerts management, and programmatic filter control for embedded analytics. That is a mixed release, but the center of gravity is obvious: make Metabase programmable, embeddable, and usable from AI and terminal workflows.

The releases page gives useful detail on the MCP direction. It says the MCP server now returns interactive charts in Claude, Cursor, or VS Code. That is not just “ask a question, get text.” It means the AI client can become a place where a user asks about data and gets a chart backed by Metabase’s permissions and semantic layer.

The 62 changelog adds lower-level evidence. Under AI, it mentions improving token consumption by loading Metabot skills on demand, fixing MCP OAuth discovery behind a TLS-terminating proxy, and supporting MCP server URL aliases. Those are not glamorous bullet points. They are the kind of operational fixes that show a feature is being run in real environments, where proxies, token bills, and URL routing break clean demos.

The June 9 AI blog supplies the product argument behind those changes. Metabase says it has shipped Metabot, MCP, Metabot in Slack, Data Studio, AI SQL and Python generation, and dashboards as code since the start of 2026. The post frames the real goal as AI that answers correctly, AI-written code that can be inspected, access controls, limits, and auditing. That is the right fear set for analytics buyers. A model that can query business data is useful only if the semantic layer, permissions, and audit trail are not a mess.

The Data Studio page supports the same story from the data-team side. It describes Data Studio as the place to transform data, centralize logic, define metrics, publish canonical tables, and track dependencies. That is less flashy than chat, but it is the foundation. If a company has three revenue definitions and six customer tables, a chatbot will not fix BI. It will just return wrong answers faster.

The AI content strategy is broader than product docs

Metabase is also publishing around AI workflows in a way that reveals how the team thinks. The Claude skill post introduces a structured Learn Metabase skill that uses Socratic teaching, memory, spaced repetition, and a user’s own data when the Metabase MCP server is connected. This is partly education, partly product adoption. If the product is getting more AI-native, the training path is getting AI-native too.

The persistent agent post is not a BI feature announcement. It describes an internal lightweight CLI agent system called pa that keeps memory and task state on disk, runs a start-of-day routine, pulls from Slack, Linear, and GitHub, tracks tasks, and can work in the background. The post claims it frees about 15% of the author’s day. Treat that as a first-person report, not a ToolVitals metric. Still, it is relevant context. Metabase is not writing about AI only as a user-facing analytics feature. It is also applying agent workflows to its own development process.

That matters because the product direction has a practical shape. Metabase is not simply putting a chat box beside dashboards. It is tying AI to MCP, semantic modeling, CLI workflows, dashboards as code, token controls, and usage auditing. The throughline is control. The product wants to let non-specialists ask questions while keeping data teams in charge of the definitions and boundaries.

There is a tension here. Metabase’s strongest historical promise is ease of use. Data Studio, MCP, AI governance, custom visualization plugins, and a CLI are more advanced surfaces. They can make Metabase more powerful, but they can also make the product feel less simple if the experience is not carefully layered. That is the maintainers’ risk: do not bury the original “ask questions without SQL” promise under infrastructure vocabulary.

What ToolVitals cannot infer

ToolVitals can see public activity. It can see stars, releases, release recency, uptime-style health signals, SSL, and source metadata when available. In this payload, it sees enough to say Metabase is highly active and scoring at the top of the tracked range.

ToolVitals cannot see whether Metabot answers are accurate on a messy warehouse. It cannot see whether a company’s semantic layer is well modeled. It cannot see support quality, enterprise renewal rates, internal security review outcomes, customer satisfaction, or how much work an admin needs to make MCP safe. It also cannot judge code quality from release counts alone.

A skeptical buyer should use the ToolVitals signal as a shortlist accelerator, not a purchasing decision. The public evidence says Metabase is a serious, active open-core analytics platform with a major AI push. Your proof should be local: connect a representative database, define canonical metrics in Data Studio, test group permissions, try MCP through the AI client your team actually uses, set token limits, inspect audit trails, and run an upgrade rehearsal with a backup of the application database.

That buyer test is especially important for embedded analytics. The official site says Metabase supports iframe embedding, a React SDK, white labeling, dynamic styling, interactive controls, and granular permissions. Those features can save engineering time, but embedded analytics lives or dies on edge cases: tenant isolation, permissions drift, theming constraints, cache behavior, and upgrade safety. ToolVitals can tell you the project is alive. It cannot tell you your multi-tenant model is safe.

For Metabase maintainers, the data says the market-facing story is strong, but the release stream could use clarity. Five releases on the same day across supported branches is healthy engineering, but buyers need to understand what changed, what is backported, what is security-sensitive, and which edition gets which AI capability. The open-core boundary is also worth keeping painfully clear. The license split is documented, but AI features, MCP, governance controls, and commercial packaging should stay easy to compare without spelunking through docs.

Recommendation

If your team wants BI that business users can operate without filing every question as a data ticket, Metabase belongs on the evaluation list. If your team also wants AI analytics, MCP access from coding tools, and a semantic layer that data teams can govern, it deserves a serious pilot rather than a casual look.

The strongest case for Metabase is not “AI dashboards are cool.” The case is that Metabase is pairing AI access with the boring controls that analytics needs: permissions, canonical metrics, dependency tracking, usage visibility, token limits, and release discipline. ToolVitals measured 19 release events in 30 days, 30 GitHub releases in 90 days, 47,723 stars, and a 100 shipping score. The first-party sources show those releases are aimed at governed AI, custom visualization, MCP, CLI, and embedded analytics work.

Pick Metabase if you want an open-core BI platform that is moving quickly toward governed AI analytics while still maintaining the old BI jobs. Be more cautious if you need a tiny, static dashboard tool, or if your team has no appetite to curate metrics and permissions. AI analytics without that foundation is just a faster way to spread bad numbers.

Sources