Graphene launches as a data analytics framework built specifically for coding agents, positioning itself as an everything-as-code stack that brings SQL governance alongside agent-friendly capabilities. The framework tackles the consistency problems that arise when teams rely on React components or Python notebooks for business intelligence, where dashboard aesthetics and metric definitions vary widely across projects.
A semantic layer that speaks SQL
Graphene SQL combines the power of standard SQL — CTEs, subqueries, window functions, set operators — with the governance concepts from traditional semantic layers. Measures and modeled joins are expressed as regular SQL expressions, so agents can invoke them without learning niche query APIs. The framework takes inspiration from Malloy, created by LookML's Lloyd Tabb and Michael Toy, but implements those ideas as good old SQL for familiarity. This means every use of a metric like "EBITDA" resolves to the same deterministic object, and metadata tags can hint chart formatting so visualizations look correct by default.
Dashboard files that enforce visual consistency
The framework introduces a dashboard file type that produces polished visuals rather than raw Python or JavaScript output. Pages support visualizations, input components for filtering, dynamic behaviors, and layout modes that distinguish monitoring-oriented dashboards from narrative-oriented notebooks. Because the markup lives in .md files version-controlled alongside code, the look and feel remains consistent across iterations.
Design goals shaped for agent ergonomics
Token efficiency guides the language design — brief syntax with minimal boilerplate. Agent ergonomics mean the entire framework is controlled via CLI, and all documentation lives inside the Graphene npm package as an agent skill. High ceilings come from ANSI SQL compliance and support for over 170 functions, while ECharts integration enables rich visualExpressions in HTML, CSS, and JavaScript.
Why coding agents choose Graphene over traditional BI
The framework appeals to teams that want to leverage their broad ecosystem of SOTA LLMs, harnesses, and skills without being locked into a single provider or proprietary tool. Business-wide context from other repos and tools becomes accessible to the agent, and end-to-end tasks can span across the stack where analytics is just one step. Version control and CI allow teams to revert agent mistakes and run tests on mission-critical dashboards. Tight iteration loops let agents validate queries, view dashboards, and iterate locally before promoting changes. Continuous agents can even self-heal codebases when schemas or metrics evolve.
Data sources and deployment
Graphene currently connects to Snowflake, BigQuery, ClickHouse, Postgres, MotherDuck, and local data via DuckDB. Adding new sources is straightforward — the team welcomes requests. Once set up, developers start the dev server with `npm exec graphene serve` and prompt their coding agent to answer data questions, build dashboards, or edit the semantic model. The CLI can compile Graphene SQL, render pages in the browser, check syntax, and print screenshots.
License and the path forward
Graphene is free under the Elastic License 2.0 for internal use cases forever. Teams building commercial applications with the framework are invited to contact the creators. The project's philosophy rejects lock-in: all files live in the user's repo, and any agent or LLM can be used. The nature of data work is expected to shift — data experts will focus on shaping the skills, models, and tools that the rest of the team uses, rather than manually building reports. Graphene Cloud is also in development as a turnkey solution to host the dashboards and reports agents build, with an MCP server and Slack bot available for pilot programs.