Graphene arrives as an everything-as-code analytics framework built for coding agents

An analytics stack whose primary user is an agent

Graphene, a data analysis framework built for coding agents, surfaced on Hacker News's front page via a Show HN post linking to its GitHub repository. The project pitches itself as an "everything-as-code" stack for SQL-based exploration, visualization and reporting, claiming that agents working in it can answer data questions and build visuals about ten times faster than the manual alternative.

According to the project's README, the framework rests on two components. The first is a semantic layer: Graphene SQL pairs conventional SQL with governed metrics and modeled joins, which the developers say produces more accurate queries than freeform SQL generation. The second is a dashboard file format designed to yield more consistent, polished visuals than agents writing chart code in Python or JavaScript from scratch.

Design choices aimed at agents

Three design goals stand out. The languages are deliberately terse, since agents pay for every token they read and write. Everything is driven through a CLI, and the entire documentation set ships as an agent skill inside the npm package rather than as a website an agent has to hunt through. And there is headroom: Graphene SQL follows ANSI SQL with more than 170 functions, while pages can express anything possible in HTML, CSS, JavaScript and the ECharts library.

The SQL dialect draws on Malloy, the query language from Lloyd Tabb and Michael Toy, the creators of LookML, but implements its ideas as ordinary SQL so models already fluent in SQL can drive it without learning a new surface syntax.

How a project is assembled

Graphene installs as a CLI through npm or an equivalent package manager. A project is either a standalone repository or a directory inside a larger codebase such as dbt, made up of .gsql files for semantic models and .md files for pages. Models declare tables, joins, dimensions and aggregating measures, and measures can compose: average order value, for example, can be defined as revenue divided by order count. Queries traverse modeled joins through a dot operator, so no explicit join clause is needed in the select. A dev server launched with npm exec graphene serve renders pages in the browser, and the CLI can also compile queries, validate syntax and capture screenshots.

Snowflake, BigQuery, ClickHouse, Postgres, MotherDuck and local files via DuckDB are the currently supported data sources, and the team says adding others is straightforward on request.

The argument against traditional BI

The README argues that agents paired with a code-based analytics stack beat conventional BI on several fronts: version control and CI, so agent mistakes can be reverted and critical dashboards can run tests; bulk refactors across many reports; and the freedom to move logic into or out of the semantic layer as needs change. Because the analytics files sit next to the rest of a company's code, an agent can also draw on broader context when deciding what to analyze.

The developers also pre-empt the suggestion that agents could simply build dashboards in React or notebooks. They concede that route works but say consistency degrades over time, with different screens quietly computing the same metric in different ways. Codifying metrics as deterministic, queryable objects is Graphene's answer, and attached metadata helps charts format values correctly by default.

Licensing and the business model

Graphene is free for internal use indefinitely, but it ships under the Elastic License 2.0, which makes it source-available rather than open source in the OSI sense; building a commercial product on top of it requires contacting the company. The commercial plan is Graphene Cloud, a managed service for hosting the dashboards and reports agents produce, alongside a hosted MCP server and Slack bot for quick questions. Anyone using the framework directly does need git and a coding agent; the browser-based experience lives in the cloud offering.

Why it matters

Most analytics tooling still assumes a human in a graphical interface. Graphene is a deliberate bet that the primary user of BI is becoming an agent working in a repository, with humans acting as reviewers. If that bet pays off, dashboards turn into software artifacts, diffable, testable and revertible, rather than documents trapped in a SaaS tool, and vendor lock-in gives way to files a company owns outright. The Elastic License caveat is worth noting for teams hoping to embed it in products they sell. More broadly, this is another data point in the shift toward agent-first tooling, and it reframes the data team's job: the README itself predicts analysts will spend less time assembling reports by hand and more time shaping the models, skills and judgment that agents depend on.