12 Best Open Source BI Tools (2026): Compared by License, Cost & AI Readiness

Open source BI used to be a budget decision. A team that couldn't justify a Tableau or Looker contract would run Metabase or Superset on a spare server, accept some rough edges, and move on, and that's still true for plenty of teams. What has changed in the last two years is who builds the analysis: a growing share of dashboards, reports and one-off analyses are now written by coding agents like Claude Code and Cursor, and that changes what you should want from an open source BI tool more than any new chart type does.

This guide covers the 12 open source BI tools we think are worth evaluating in 2026, with the license, GitHub activity, cloud pricing, self-hosting needs and AI readiness of each.

The quick answer: which open source BI tool fits which team

• anfra fits teams that want coding agents to build custom analytics apps on governed metrics.
• Metabase fits non-technical teams that want point-and-click questions and dashboards with the least setup.
• Apache Superset fits data teams that need a full dashboarding platform under the Apache license.
• Lightdash fits dbt teams that want BI defined in the same repository as their models.
• Redash fits SQL-first analysts who mostly write queries and share the results.
• Evidence fits teams that want reports written as code (Markdown plus SQL) and published as a static site.
• Grafana fits operational and time-series monitoring more than business reporting.
• Cube fits teams that need a standalone semantic layer feeding several front ends.
• Rill fits teams that want very fast dashboards on ClickHouse or DuckDB, defined as code.
• WrenAI fits teams that want an open source text-to-SQL layer for natural-language questions.
• Knowage fits organizations that need a traditional, all-in-one BI suite under a copyleft license.
• Eclipse BIRT fits Java teams that need paginated, printable reports embedded in an application.

Comparison table

GitHub figures are a snapshot as of October 8, 2026, and cloud prices were checked on the same date.

#	Tool	License	GitHub stars	Latest release	Self-host footprint	Managed cloud and starting price	Semantic layer	AI / agent support	Best for
1	anfra	Apache 2.0	New (launched 2026)	v0.4.3 (Oct 2026)	Single Go binary	anfra Cloud (waitlist); price not published	Yes (AMQL, built in)	Built for coding agents: MCP server and agent skill	Agent-built custom BI apps
2	Metabase	AGPL-3.0 (OSS edition); commercial license for enterprise/	~49.6k	v0.64.1 (Oct 2026)	JVM app + Postgres/MySQL app DB	Metabase Cloud: Starter $100/mo incl. 5 users, +$6/user	Light (models, metrics)	Metabot, agent API	Non-technical self-serve
3	Apache Superset	Apache 2.0	~75.1k	6.1.0 (May 2026)	Python app + Postgres + Redis + Celery workers	Preset: free up to 5 users; Professional $20/user/mo	Light (virtual datasets, metrics)	Limited in core	Full dashboarding platform
4	Lightdash	MIT, with a commercial ee/ directory	~6.2k	2.475.2 (Oct 2026)	Node app + Postgres, reads your dbt project	Lightdash Cloud: Pro $3,000/mo flat	Yes (dbt-based)	AI agents, MCP, CLI	dbt-centric teams
5	Redash	BSD-2-Clause	~28.8k	v26.9.0 (Sep 2026)	Python app + Postgres + Redis	No official cloud; third-party hosting from ~$11/mo	No	None built in	SQL-first query sharing
6	Evidence	MIT	~7.0k	40.1.8 (Feb 2026)	Static site build (Node)	Evidence: Team $2,500/mo flat	No (SQL in Markdown)	Works with Claude Code and Cursor	Reports as code
7	Grafana	AGPL-3.0	~77.1k	v13.2.3 (Sep 2026)	Single Go binary + SQLite/Postgres	Grafana Cloud: free tier; Pro $19/mo + usage	No	Grafana Assistant	Ops and time-series monitoring
8	Cube (Cube Core)	Apache 2.0 (some packages MIT)	~21.0k	v1.8.1 (Oct 2026)	Node API + Cube Store (Rust)	Cube Cloud: free tier; Starter $40/developer/mo + compute	Yes (standalone)	APIs for agents; most AI features in Cloud	Headless semantic layer
9	Rill	Apache 2.0	~2.9k	v0.90.4 (Oct 2026)	Single Go binary with embedded DuckDB	Rill Cloud: Starter $20/seat/mo + compute	Yes (metrics views in YAML)	Scaffolds Claude/Cursor instructions and MCP	Fast OLAP dashboards as code
10	WrenAI	Apache 2.0 for core and SDK; docs CC BY 4.0	~17.8k	wren-v0.15.0 (Sep 2026)	Self-hosted engine	Wren AI Cloud: free tier; Essential $179/mo (annual)	Yes (semantic model, MDL)	Text-to-SQL for agents	Natural-language querying
11	Knowage	AGPL-3.0	~450	v9.1.0 (Oct 2026)	Java (Tomcat) + MariaDB/MySQL	Enterprise edition; price not published	Yes (business models)	Limited	Traditional all-in-one BI suite
12	Eclipse BIRT	EPL-2.0	~550	4.25.0 (Sep 2026)	Java library / runtime	No hosted version	No	None	Embedded paginated reports

How AI changes what "open source BI" should mean

Every BI tool makes two promises: anyone can ask the question they actually have, and the answer means what the organization thinks it means.

The first promise is about freedom (can a person get an answer without waiting for an analyst?) and the second is about trust (can anyone vouch for the number?). The two pull against each other, and each generation of BI tools has traded some of one for more of the other.

IT-built reports and OLAP cubes gave organizations trust by limiting both what could be asked and what it meant, which is why every request went through an analyst backlog. Desktop self-service tools went the other way: anyone could drag fields onto a canvas, and logic forked per workbook until two teams brought two revenue numbers to the same meeting. Semantic-layer BI (Looker, dbt's semantic layer, Holistics and, in open source, Lightdash, Cube and a lighter version in Superset) moved metric definitions back into a central model while leaving people free to explore inside it.

AI agents give people more freedom than any earlier generation of tools, anda most open source BI tools were designed before agents existed. When agent queries the warehouse directly, trust often erodes in three ways:

1. Correctness. Some errors are about definitions, where the agent picks a reasonable meaning of "active customer" that is the wrong one for your company. Others are about computation and are harder to spot, because the output still looks right: averaging a precomputed conversion rate across regions instead of dividing total conversions by total visits, or summing an order-level shipping fee after joining to order line items, which counts each fee once per line.
2. Consistency. Three people asking three agents for "active users" over six months get three reasonable and different answers. Self-service BI had the same problem, but its copies of the logic lived in SQL files and dashboards that someone could find and compare, whereas agent-generated logic lives in chat transcripts and generated HTML that nobody reviews.
3. Collaboration. People trust an answer partly because others can check it and build on it. A chat transcript keeps the reasoning but is private and hard to revisit, and a pasted chart is easy to share but has the queries and definitions removed.

What this means for evaluating open source BI tools in 2026

The usual checklist (chart types, connectors, dashboard quality, ease of install) still matters, but it doesn't tell you what happens when an agent does the building. We added five criteria that follow from the problems above:

1. Binding definitions. Metrics live in a layer the agent has to use, rather than in documentation it can read and ignore. A tool with no semantic layer leaves every agent to define revenue again for each question.
2. Valid composition. The engine knows each metric's grain and whether it can be summed, so averaging a ratio or summing across a join that duplicates rows returns an error instead of a wrong number that looks fine.
3. Expressiveness. The semantic layer can express cohorts, period-over-period comparisons and ratios at the right grain. When it can't, people and agents fall back to raw SQL, and the definitions in that SQL are no longer governed (we call this semantic leakage).
4. Inspectability. Every number traces back to a readable query and the definitions behind it, so a person can check an agent's work in a minute or two.
5. Durable, editable output. The agent produces a structured object that others can open, change and build on, rather than a screenshot or a page of generated HTML.

Most of the established BI tools on this list do well on the usual checklist and partly meet these five, which is reasonable given when they were designed. The tools that do best on the new criteria are the ones with a semantic layer that sits in the query path: anfra, Cube, Lightdash, Rill and, for natural-language questions, WrenAI.

How we evaluated

We ran each tool against a sample Postgres and DuckDB dataset in September and October 2026. We checked every license file, pulled GitHub stars and release dates from the GitHub API on October 8, 2026, and read each vendor's pricing page on the same day.

We counted a tool as open source only when its core is published under an OSI-approved license, which is why free but proprietary tools such as Tableau Public and Power BI Desktop are left out, and why open-core products are labeled with the parts that are commercial.

We scored each tool on ten criteria, five from the usual BI checklist and five from the AI section above:

Usual BI criteria

1. Setup and maintenance effort. How long it takes to get from download to a working dashboard, how many services a production deployment needs, and how painful upgrades between major versions are.
2. Analysis and visualization. The range of chart types, dashboard features (filters, drill-downs, cross-filtering) and whether non-technical users can build questions without SQL.
3. Data source coverage. Which warehouses and databases it connects to, and how well it handles large tables.
4. Permissions and embedding in the open source edition. Whether SSO, role-based access, row-level security and white-label embedding are included or sit in a paid tier.
5. License, project health and cost. The license and what it means for embedding, how often releases ship, the size of the community, and what the managed cloud version costs.

AI-era criteria

1. Binding definitions. Whether metrics are defined in a layer that queries must go through.
2. Valid composition. Whether the engine blocks invalid aggregations such as averaging a ratio or summing across a fan-out join.
3. Expressiveness. Whether cohorts, period comparisons and ratios can be expressed without dropping to raw SQL.
4. Inspectability. Whether a person can see the query and definitions behind any number.
5. Durable, editable output. Whether what an agent produces can be opened, edited and reused by others.

Disclosure: anfra is built by the team at Holistics, which publishes this blog. Holistics itself is a commercial BI platform and is left off this list because it isn't open source.

In each tool profile below, every criterion gets a star rating: ★★★ is strong, ★★☆ is partial and ★☆☆ is weak.

The 12 best open source BI tools

1. anfra

A funnel analysis app built by a coding agent with anfra (October 2026). Source: anfra on GitHub.

anfra is an open source engine that lets coding agents build custom analytics apps on a governed semantic layer. The agent writes the app's HTML, JavaScript and charts, but every number comes from the semantic layer, not from SQL the agent improvised. It has three parts:

• A JavaScript SDK that apps use to request data by metric name, never raw SQL.
• A server (anfra serve) that runs those requests on your warehouse and exposes an MCP endpoint for agents.
• AMQL, the semantic layer Holistics uses in production, with models and metrics defined as code in your repository.

Once it's set up, building an app takes a single prompt. Ask your agent to "build a custom dashboard from these data models" and it can produce a working BI app in one shot.

How anfra works.

What sets it apart from the other tools on this list:

• Fully custom front ends. There are no fixed chart types or dashboard layouts. The agent builds whatever interface the analysis needs, from a KPI page to a game-like experience, on the same governed metrics.
• No metric drift. Apps can only request metrics defined in the semantic layer and can't change the model, so every app shows the same revenue number.
• Every number can be checked. An Inspect panel on each chart shows the query, the metric definitions and the SQL that ran.
• Advanced analysis built in. Funnels, cohort retention and period-over-period comparisons live in the semantic layer, so apps don't each carry their own SQL for them.
• Models reviewed like code. If no models exist yet, the agent drafts them and your data team reviews them in a pull request.
• Runs on your infrastructure. It ships as a single binary and connects to Snowflake, BigQuery, Databricks, Redshift, PostgreSQL, ClickHouse, DuckDB and more.

Custom, even playful, data experiences (here, a BI tool turned into a video game) without giving up governance.

• Limitations: anfra is the youngest project on this list (v0.4.x, first released in 2026), so it has a smaller community and faster-changing APIs than Metabase or Superset. AMQL is the only supported semantic layer for now; support for dbt, Cube, Snowflake and Databricks semantic layers is planned.
• Best for: Data teams whose analysts already build with coding agents and want governed apps that others can check, not one-off dashboards built on ad hoc SQL.

How anfra scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★★	Single binary; anfra setup wires up Claude Code or Cursor in one step.
Analysis and visualization	★★☆	Any chart or layout an agent can build, but no point-and-click builder for non-technical users.
Data source coverage	★★★	Snowflake, BigQuery, Databricks, Redshift, PostgreSQL, ClickHouse, DuckDB and others.
Permissions and embedding (OSS)	★☆☆	The open source server has no login, permissions or row-level security; those are in anfra Cloud.
License, health and cost	★★☆	Apache 2.0 and free, but the youngest project here, with a small community and fast-changing APIs.
Binding definitions	★★★	Apps can only request metrics defined in AMQL and can't change the model.
Valid composition	★★★	Metrics carry their aggregation and grain in AMQL, so ratios and fan-out joins are computed at the right level.
Expressiveness	★★★	Funnels, cohort retention and period-over-period comparisons are built into the semantic layer.
Inspectability	★★★	An Inspect panel on every chart shows the query, metric definitions and the SQL that ran.
Durable, editable output	★★★	Apps and models are code in your repository, so others can review, edit and extend them.

2. Metabase

The current Metabase product UI (August 2026). Source: Metabase on GitHub.

Metabase is the open source BI tool most people try first. A non-technical user can go from a fresh install to a working question in minutes using a visual query builder that doesn't require SQL. The open source edition covers questions, dashboards, models, alerts and subscriptions, and users who know SQL can switch to a native editor when the visual builder isn't enough.

Features such as advanced permissions, SSO beyond the basics and interactive embedding without Metabase branding are in paid editions. Metabase has added Metabot (an AI assistant that answers questions and helps write queries) and an agent API. Its modeling layer is lighter than a dedicated semantic layer, so agents and users can still query raw tables and write their own definitions.

• Hosting: JVM application plus a Postgres or MySQL application database, or Metabase Cloud.
• Cloud pricing: Metabase Cloud has no free plan (there is a 14-day trial). Starter costs $100/month including 5 users, plus $6 per additional user. Pro costs $575/month plus $12 per user and adds SSO, row- and column-level security, embedding and white-labeling. Enterprise starts at $20,000/year.
• Best for: Teams that want the easiest start for business users and can accept a lighter metric layer.

How Metabase scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★★	The fastest path from install to first dashboard on this list; one JVM app plus an app database.
Analysis and visualization	★★★	Visual query builder, dashboards with filters and drill-through, and a native SQL editor.
Data source coverage	★★★	Connects to all major warehouses and databases, with community drivers for more.
Permissions and embedding (OSS)	★★☆	Basic groups and permissions are free; SSO beyond the basics, row-level security and white-label embedding are paid.
License, health and cost	★★☆	Large community and frequent releases, but AGPL-3.0 plus a commercial enterprise/ directory.
Binding definitions	★★☆	Models and metrics exist, but users and agents can still query raw tables and define their own.
Valid composition	★☆☆	No grain awareness; nothing stops averaging a ratio or summing across a join that duplicates rows.
Expressiveness	★★☆	Custom expressions cover common cases; cohorts and complex ratios usually need SQL.
Inspectability	★★★	Any question built in the visual builder can show its generated SQL.
Durable, editable output	★★★	Questions, models and dashboards are saved objects others can open, duplicate and edit.

3. Apache Superset

A dashboard in Apache Superset 6 (May 2026). Source: superset.apache.org.

Apache Superset is the most complete open source dashboarding platform, with more than 40 chart types, a SQL editor (SQL Lab), a light semantic layer of virtual datasets and metrics, and role-based access control in the open source build.

The trade-off is that Superset takes more work to run. A production deployment usually includes the web app, a metadata database, Redis and Celery workers for async queries and caching, and upgrades between major versions need planning. Preset, founded by Superset's creator, sells a managed version, which is the usual option for teams that want Superset without operating it. AI features are limited in core Superset, and its semantic layer is thin enough that agents working with it will mostly write their own SQL.

• Hosting: Python application with Postgres, Redis and Celery, or Preset.
• Cloud pricing (Preset): Preset is free for up to 5 users. Professional costs $20 per user per month on an annual plan ($25 month to month). Embedded dashboards cost $500/month per 50 viewers, and SSO, SCIM, audit logs and Preset's AI features are on the custom-priced Enterprise plan.
• Best for: Data teams that need a full dashboarding platform and have the engineering time to operate it.

How Apache Superset scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★☆☆	Production needs the app, a metadata database, Redis and Celery workers; major upgrades need planning.
Analysis and visualization	★★★	More than 40 chart types, cross-filtering, drill-downs and SQL Lab.
Data source coverage	★★★	Connects to almost any database with a SQLAlchemy driver.
Permissions and embedding (OSS)	★★★	Role-based access, row-level security and an embedding SDK are in the open source build, though they take configuration.
License, health and cost	★★★	Apache 2.0 under the Apache Software Foundation, with a large, active community.
Binding definitions	★☆☆	Metrics live on individual virtual datasets, and SQL Lab lets anyone bypass them.
Valid composition	★☆☆	No grain awareness across datasets; a metric is a SQL expression applied to whatever rows it gets.
Expressiveness	★★☆	Time comparison and rolling windows are chart options; cohorts and governed ratios need SQL.
Inspectability	★★★	Every chart can show the query it ran.
Durable, editable output	★★★	Charts and dashboards are saved objects and can be exported as YAML.

4. Lightdash

The Lightdash platform (July 2026). Source: Lightdash on GitHub.

Lightdash reads your dbt project and turns the models, dimensions and metrics defined in dbt YAML into a BI layer people can explore, so the definitions live in the same repository and review process as your transformations. It now describes itself as open source "agentic BI", with AI agents, MCP support and a CLI for previewing and shipping changes through Git and CI.

Lightdash does well on binding definitions because metrics come from dbt, and less well for teams without dbt, since the whole product assumes it.

• Hosting: Node application plus Postgres, connected to your dbt project, or Lightdash Cloud.
• Cloud pricing: Lightdash Cloud has no free plan (there is a 21-day trial). Cloud Pro costs $3,000/month flat with unlimited users and includes AI agents and MCP. Embedding is an add-on, and SAML, SCIM and custom roles are on Enterprise.
• Best for: dbt-centric teams that want BI defined alongside their models.

How Lightdash scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★☆	A Node app plus Postgres, easy if you already run dbt and a hurdle if you don't.
Analysis and visualization	★★☆	A clean explore interface and solid chart set, narrower than Superset's.
Data source coverage	★★☆	Limited to warehouses dbt supports, which covers the major cloud warehouses.
Permissions and embedding (OSS)	★★☆	Project roles are open source; embedding and SAML/SCIM are in the commercial ee/ code or Enterprise.
License, health and cost	★★★	MIT outside ee/, with very frequent releases.
Binding definitions	★★★	Explores are built from metrics defined in dbt YAML, reviewed with your transformations.
Valid composition	★★☆	Joins and metric types are declared in dbt, but protection against fan-out joins depends on how they're modeled.
Expressiveness	★★☆	Custom metrics and table calculations cover period comparisons; cohorts usually need a dbt model.
Inspectability	★★★	Every chart shows its compiled SQL.
Durable, editable output	★★★	Charts and dashboards can be downloaded, edited and uploaded as code through the CLI.

5. Redash

Redash's query editor, charts and dashboards. Redash has published no new official screenshots since about 2019, and the UI has changed little. Source: Redash on GitHub.

Redash is SQL-first: analysts write queries, turn the results into charts, and combine them into dashboards that others can view or refresh. It connects to a long list of data sources, and has a REST API that covers everything in the UI. After Databricks acquired Redash in 2020, the hosted service shut down and development slowed for a while, but the community project is active again, with a v26.9.0 release in September 2026.

Redash has no semantic layer and no built-in AI, so every query has its own definitions. That makes it a good fit for a small analyst team sharing SQL and a poor fit for governed self-service or agent-built analysis.

• Hosting: Python application with Postgres and Redis.
• Cloud pricing: Redash has had no official managed cloud since the hosted service shut down after the Databricks acquisition. Third-party hosts such as Elestio run managed Redash from about $11/month.
• Best for: SQL-fluent analysts who mainly need to write, chart and share queries.

How Redash scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★☆	A Python app with Postgres and Redis, simple with the official Docker setup.
Analysis and visualization	★★☆	Basic charts and dashboards on top of SQL; nothing for users who don't write SQL.
Data source coverage	★★★	One of the longest data source lists here, including APIs and NoSQL stores.
Permissions and embedding (OSS)	★★☆	Groups, data source permissions and SAML/Google login are free; embedding is limited to shared links.
License, health and cost	★★☆	Permissive BSD-2-Clause, and development has picked up again after a slow period.
Binding definitions	★☆☆	No semantic layer; every query defines its own metrics.
Valid composition	★☆☆	Nothing checks how a query aggregates.
Expressiveness	★☆☆	Anything SQL can express, but none of it is governed.
Inspectability	★★★	Every chart is a visible, readable query.
Durable, editable output	★★★	Queries are saved, forkable and refreshable by others.

6. Evidence

Evidence Core, the open source edition (August 2026). Source: Evidence blog.

Evidence treats reports as code: you write Markdown with embedded SQL queries and chart components, and Evidence builds a static website from them. Reports are version-controlled, reviewable and cheap to host, which suits analysts who already work in a code editor. Because each report is a Markdown file, coding agents such as Claude Code or Cursor can write and edit reports directly.

The trade-off is that SQL lives in each report, so two reports can define the same metric differently unless the team adds review rules or an upstream semantic layer. Its latest tagged release is from February 2026, though the repository still receives regular commits.

• Hosting: Static site build (Node) that you can host anywhere.
• Cloud pricing: Evidence has no free cloud plan (there is a 30-day trial). Team costs $2,500/month flat with unlimited users and 20,000 AI credits. SSO, SCIM, row-level access, embedding and white-labeling are on the custom-priced Enterprise plan.
• Best for: Analysts who want version-controlled, publishable reports written as code.

How Evidence scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★★	npm install and a static build you can host anywhere.
Analysis and visualization	★★☆	A good library of chart components, but every report is written in code.
Data source coverage	★★★	Connectors for the major warehouses, databases, DuckDB and files.
Permissions and embedding (OSS)	★☆☆	A static site has whatever access control your host gives it; SSO and row-level access are Enterprise.
License, health and cost	★★☆	MIT licensed, but the last tagged release was February 2026.
Binding definitions	★☆☆	SQL lives in each report, so two reports can define the same metric differently.
Valid composition	★☆☆	Nothing checks how a query aggregates.
Expressiveness	★★☆	Anything SQL can express, without governance.
Inspectability	★★★	The SQL behind every chart sits in the same Markdown file.
Durable, editable output	★★★	Reports are Markdown files in Git that people and agents can edit and review.

7. Grafana

A Grafana dashboard with AI features (January 2026). Some features shown may be specific to Grafana Cloud. Source: grafana.com.

Grafana is the most widely used open source tool for operational dashboards over metrics, logs and traces, with strong support for Prometheus, Loki, Elasticsearch and dozens of time-series and SQL sources. It appears in BI searches because it can query SQL databases and build good-looking dashboards, but its strengths (real-time panels, alerting, observability data) are different from what business reporting needs, such as drill-downs across customers and products or governed revenue metrics.

Grafana Labs ships an AI assistant in its cloud product.

• Hosting: Single Go binary with SQLite or Postgres, or Grafana Cloud.
• Cloud pricing: Grafana Cloud has a free tier (3 users, 10,000 metric series, 50 GB of logs). Pro starts at $19/month plus usage, including $8 per active user, and Enterprise has a $25,000/year minimum.
• Best for: Engineering and operations monitoring, used alongside a business BI tool rather than as one.

How Grafana scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★★	A single Go binary with SQLite, or Postgres for production.
Analysis and visualization	★★☆	Excellent time-series panels and alerting; weak for business drill-downs across customers and products.
Data source coverage	★★★	Dozens of observability and SQL sources.
Permissions and embedding (OSS)	★★☆	Teams, folder permissions and OAuth/LDAP are free; data source permissions and reporting are Enterprise.
License, health and cost	★★☆	Very active and widely used, but AGPL-3.0.
Binding definitions	★☆☆	No semantic layer; each panel holds its own query.
Valid composition	★☆☆	Nothing checks how a query aggregates.
Expressiveness	★☆☆	Transformations help with time series, but business metrics need hand-written queries.
Inspectability	★★★	The query inspector shows the request and response behind every panel.
Durable, editable output	★★★	Dashboards are JSON that can be versioned and provisioned as code.

8. Cube

How Cube Core sits between your data sources and downstream tools (September 2026). Cube Core has no UI of its own; the dashboards in the video below are Cube Cloud. Source: Cube on GitHub.

Cube Core is an open source, standalone semantic layer. You define metrics, dimensions, joins and access rules once in code, and Cube exposes them through SQL, REST and GraphQL APIs to whatever sits downstream: BI tools, embedded analytics in your own product, or AI agents. It includes a caching layer (Cube Store) for pre-aggregations, which is often the main reason teams adopt it.

Cube has no BI front end of its own, so you pair it with Superset, Metabase, a custom app or an agent. Many of Cube's AI and visualization features are in Cube Cloud only.

• Hosting: Node API plus Cube Store, or Cube Cloud.
• Cloud pricing: Cube Cloud has a free tier (1,000 API requests per day). Starter costs $40 per developer per month plus compute, Premium costs $80 per developer and adds viewer roles and embedding, and SSO, audit logs and bring-your-own-LLM are on Enterprise.
• Best for: Teams that want one semantic layer feeding several front ends, especially for embedded analytics.

How Cube scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★☆	A Node API plus Cube Store, and you still need a front end.
Analysis and visualization	★☆☆	No end-user interface of its own; the Playground is for developers.
Data source coverage	★★★	Connects to all major warehouses and many databases.
Permissions and embedding (OSS)	★★☆	Row-level access rules and security context are in Cube Core; users, SSO and dashboards come from your front end.
License, health and cost	★★★	Apache 2.0, active, with a large community.
Binding definitions	★★★	Front ends and agents query metrics through Cube's APIs rather than raw tables.
Valid composition	★★★	Measures declare their type and joins declare their relationship, which Cube uses to avoid double-counting across fan-out joins.
Expressiveness	★★☆	Rolling windows and multi-stage measures cover period comparisons; cohorts take extra modeling.
Inspectability	★★☆	Generated SQL is visible in the Playground and logs, but end users only see what their front end shows.
Durable, editable output	★★☆	The model is code; whether results are durable depends on the front end you pair it with.

9. Rill

A Rill Explore dashboard with several measures selected (July 2025). Source: Rill docs.

Rill builds very fast exploratory dashboards on OLAP engines, with DuckDB embedded for local work and ClickHouse for production scale. Projects are defined as code in YAML and SQL, including metrics views that act as a semantic layer, and Rill's CLI sets up Claude Code instructions and an MCP configuration so agents can create projects, dashboards and security policies.

• Hosting: Single Go binary, or Rill Cloud.
• Cloud pricing: Rill Cloud has no free plan (new accounts get $250 of usage credit). Starter costs $20 per seat per month plus compute and storage, embedding requires Growth ($500/month per 100 seats), and SAML/Okta SSO is on Enterprise.
• Best for: Teams that need dashboards that respond in under a second over large event data, defined as code.

How Rill scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★★	A single binary with DuckDB embedded.
Analysis and visualization	★★☆	Fast slice-and-dice explore dashboards; less suited to bespoke report layouts.
Data source coverage	★★☆	Built around OLAP engines (DuckDB, ClickHouse, Druid, Pinot), ingesting from warehouses and files.
Permissions and embedding (OSS)	★★☆	Row-level security policies are defined in YAML; sharing and SSO are in Rill Cloud.
License, health and cost	★★☆	Apache 2.0 and frequent releases, but a smaller community.
Binding definitions	★★★	Dashboards are built on metrics views, not ad hoc queries.
Valid composition	★★☆	Metrics views sit on one modeled table, which sidesteps fan-out joins rather than checking them.
Expressiveness	★★☆	Time comparisons are built in; cohorts and multi-table ratios need upstream modeling.
Inspectability	★★☆	Measures are readable SQL expressions in YAML, though dashboards don't show the full query by default.
Durable, editable output	★★★	Projects, metrics and dashboards are all YAML and SQL files.

10. WrenAI

A Wren AI GenBI app (October 2026). Source: getwren.ai.

WrenAI is an open source "GenBI" engine: you define a semantic model of your data, and agents or users ask questions in natural language that WrenAI turns into SQL across more than 22 data sources. It can also generate charts and dashboards, and it keeps business definitions in Git.

WrenAI is the most direct open source option for "chat with your data", and its semantic model gives it better grounding than plain text-to-SQL. The output is still SQL generated for each question, though, so check the generated queries before trusting complex metrics.

• Hosting: Self-hosted engine, or Wren AI Cloud.
• Cloud pricing: Wren AI Cloud has a free tier with 20 credits per month. Essential costs $179/month on an annual plan ($259 month to month) and is credit-based, Enterprise Cloud ($559/month) adds row- and column-level security, MCP and audit logs, and SSO, SCIM and on-premises deployment are on the custom-priced Enterprise Plus plan.
• Best for: Teams that want natural-language querying over a defined semantic model.

How WrenAI scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★☆	Docker-based setup plus an LLM provider, and the scope of the open source edition is shifting.
Analysis and visualization	★★☆	Generates charts and simple dashboards from questions.
Data source coverage	★★★	More than 22 data sources.
Permissions and embedding (OSS)	★☆☆	Row- and column-level security, SSO and audit logs are in the paid cloud plans.
License, health and cost	★★☆	Apache 2.0 and active, but with the open source scope in flux.
Binding definitions	★★☆	The semantic model grounds every question, but the LLM still writes new SQL each time.
Valid composition	★☆☆	Relationships are modeled, but the LLM chooses how to aggregate.
Expressiveness	★★☆	Anything SQL can express, with only the model as guidance.
Inspectability	★★★	Shows the generated SQL and the reasoning steps for each answer.
Durable, editable output	★★☆	Answers can be saved and pinned to dashboards, but most output lives in question threads.

11. Knowage

A Knowage interactive dashboard (April 2026). Source: knowage-suite.com.

Knowage (the successor to SpagoBI) is a traditional, all-in-one BI suite covering reporting, dashboards, OLAP, a business-model layer and data preparation. It is maintained by Engineering Group, which sells an enterprise edition. It still ships releases regularly (v9.1.0 in October 2026), though its community is much smaller than Metabase's or Superset's.

• Hosting: Java application on Tomcat with MariaDB or MySQL.
• Cloud pricing: Knowage has no managed cloud. The Community Edition is free, and the Enterprise Edition (AI assistant, scheduler, multi-tenancy, audit, data preparation) is sold by Engineering Group with prices available on request.
• Best for: Organizations, often in the public sector or in Europe, that want a classic BI suite under a copyleft license.

How Knowage scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★☆☆	A Java app on Tomcat with MariaDB or MySQL, and heavier to run than most tools here.
Analysis and visualization	★★☆	Reporting, OLAP and dashboards in one suite, with a dated interface.
Data source coverage	★★★	JDBC sources plus files and NoSQL stores.
Permissions and embedding (OSS)	★★☆	Role-based access and profile-based data restrictions in the Community Edition.
License, health and cost	★☆☆	AGPL-3.0 with a small community, though releases are regular.
Binding definitions	★★☆	Business models define shared entities, but users can still build their own datasets.
Valid composition	★☆☆	No checks on aggregation across joins that we could find.
Expressiveness	★★☆	OLAP cubes support period and hierarchical analysis.
Inspectability	★★☆	Dataset queries are visible to developers but not to most report viewers.
Durable, editable output	★★★	Reports and dashboards are saved documents others can open and edit.

12. Eclipse BIRT

The BIRT Report Designer in Eclipse (2021; the designer has changed little since). Source: eclipse-birt.github.io.

Eclipse BIRT is a reporting engine rather than a self-service BI tool: Java developers design paginated, print-ready reports (invoices, statements, regulatory reports) and embed the runtime in an application. It still releases regularly (4.25.0 in September 2026). JasperReports is the other long-standing option in this category and is worth comparing if your team is choosing a Java reporting library.

• Hosting: Java library and runtime inside your application.
• Cloud pricing: BIRT has no hosted cloud version. Commercial support is available from consultancies such as Innovent Solutions and through OpenText's BI and reporting products, with prices available on request.
• Best for: Java teams that need printable, paginated reports embedded in a product.

How Eclipse BIRT scores against our criteria

Criterion	Rating	Why
Setup and maintenance	★★☆	A library you embed in a Java app, which is developer work rather than an install.
Analysis and visualization	★★☆	Strong paginated, print-ready layouts; no exploratory or self-service analysis.
Data source coverage	★★☆	JDBC, XML, CSV and scripted data sources.
Permissions and embedding (OSS)	★☆☆	Inherits whatever the host application provides.
License, health and cost	★★☆	EPL-2.0 with regular releases and a small community.
Binding definitions	★☆☆	No semantic layer.
Valid composition	★☆☆	Nothing checks how a query aggregates.
Expressiveness	★★☆	Computed columns and aggregations inside report designs.
Inspectability	★★☆	Queries live in the report design file, visible to developers.
Durable, editable output	★★★	Report designs are XML files that can be versioned and edited.

Free, but not open source BI Tools

Tableau Public and Power BI Desktop are free to use but proprietary, and Tableau Public publishes every workbook publicly. FineReport, which often appears on lists like this one, is commercial software with a free trial. We left all three off the ranked list for that reason.

What the next generation of AI-native open source BI looks like

Most of the tools on this list were designed for people building dashboards by hand, and they are adding AI as a feature on top. The more useful question is what an open source BI tool looks like when it is designed for agents from the start.

Our view is that it should control what metrics mean and leave people free in how they ask and build. Each earlier generation of BI gained freedom by giving up trust, or gained trust by giving up freedom. The combination nobody has delivered yet is open interaction with governed meaning: people and agents can ask anything, build any layout and go into any level of detail, while every number goes through definitions the agent can't bypass. In practice, the engine has to enforce the five properties described earlier (binding definitions, valid composition, expressiveness, inspectability and durable output) rather than relying on instructions in a prompt.

The usual objection is that someone has to build the semantic layer, and that modeling work stalled many past projects. AI changes this too, because an agent can draft models as easily as it drafts queries. The data team's job then shifts from writing definitions to reviewing them in a pull request, and the model can grow based on demand, starting with the concepts people ask about most.

Several projects already show parts of this approach: Lightdash and Rill expose their modeling layers to agents through MCP, WrenAI grounds natural-language questions in a semantic model, Cube offers a semantic layer any agent can call, and anfra puts the semantic layer between the agent's app code and the warehouse so the app can only request defined metrics. The open question for the next few years is which of these approaches still works when hundreds of agent-built apps and analyses pile up inside one company, and whether two people asking the same question six months apart still get the same answer.

How to choose an open source BI tool

The right choice depends mostly on who builds the analysis and how much governance you need:

1. A business team that wants self-service with minimal setup should start with Metabase, and a team that requires the Apache license can use Superset (or Preset) instead.
2. A dbt team that wants metrics and dashboards defined in one place should start with Lightdash.
3. A team whose analysts already build with Claude Code or Cursor should look at anfra first, with Rill as an alternative for OLAP-style dashboards and Evidence for static reports.
4. A product team embedding analytics in its own application usually wants a permissively licensed semantic layer (Cube, or anfra for agent-built front ends) and should check AGPL obligations before choosing Metabase or Grafana.
5. A team that mainly wants natural-language questions can evaluate WrenAI, checking the generated SQL against a set of questions with known answers before rolling it out.
6. An engineering or operations team monitoring systems should use Grafana.
7. A Java team that needs printable reports should choose between BIRT and JasperReports.

Before committing, it's worth running one real question through each shortlisted tool, ideally one with a ratio metric and a join that can duplicate rows, and checking three things: whether the number is right, whether a colleague can see how it was computed, and whether asking again next week gives the same answer.

FAQ

What is the best open source BI tool?

The best tool depends on your team. Metabase is the easiest for non-technical users, Apache Superset is the most complete for data teams, Lightdash fits dbt users, and anfra fits teams building analytics with coding agents on a governed semantic layer.

Are open source BI tools really free?

The software is free to download and run, but you pay for infrastructure, for engineering time spent on upgrades and maintenance, and often for a paid edition once you need SSO, detailed permissions, row-level security or white-label embedding. Most of the tools on this list also sell a managed cloud version, with prices listed in each section above.

Is Metabase really open source?

Metabase's open source edition is licensed under AGPL-3.0, while the code in its enterprise/ directory is under a commercial license. Features such as advanced permissions and some embedding options require a paid plan.

What is the best open source alternative to Power BI or Tableau?

Apache Superset is the closest in range of charts and dashboards, and Metabase is the closest in ease of use for business users. Neither replicates Power BI's DAX modeling or Tableau's visual analysis exactly, so test with your own reports before migrating.

Which open source BI tools are on GitHub?

All 12 tools in this guide publish their source on GitHub: anfra (holistics/anfra), Metabase, Apache Superset, Lightdash, Redash, Evidence, Grafana, Cube, Rill, WrenAI, Knowage and Eclipse BIRT. Stars and latest release dates are in the comparison table above.

What is a self-hosted BI tool?

A self-hosted BI tool runs on infrastructure you control (a VM, Kubernetes or your own cloud account) rather than the vendor's cloud, which keeps data access inside your network. Every tool on this list can be self-hosted.

Can AI agents use open source BI tools?

Yes, and how safely depends on whether the tool has a semantic layer the agent must query through. anfra, Rill and Lightdash expose their models to agents through MCP, Cube exposes a semantic layer over APIs, and WrenAI converts natural-language questions into SQL based on a semantic model. Tools without a semantic layer leave the agent to write its own SQL and definitions.

Do I need a semantic layer for open source BI?

A small team writing its own SQL can manage without one. Once several people, or several AI agents, are answering questions about the same metrics, a semantic layer is what keeps "revenue" meaning the same thing everywhere.