Microsoft Research, in partnership with Renmin University of China’s IDEAS Lab, has open-sourced an intermediate language for data visualization designed to streamline how artificial intelligence agents generate graphical charts.
The new specification, named Flint, functions as a translation layer that allows artificial intelligence agents to generate renderable charts for popular visualization libraries like Vega-Lite, ECharts, and Chart.js using single-sentence natural language commands. The project has been made publicly available on GitHub.
Existing visualization libraries offer highly complex configuration frameworks. When artificial intelligence developers use streamlined presets, the resulting charts often fail to match the underlying data appropriately. Conversely, when developers task large language models with outputting complete, raw configurations, the excessive parameters frequently trigger architectural conflicts, omissions, or logical inconsistencies.
To bypass these hurdles, Flint decouples the chart's structural intent from the specific implementation rules of the underlying visualization library, reducing the configuration complexity handled by the artificial intelligence.
The language splits its specification into two primary domains, data definition and chart definition. The data layer annotates the specific semantic types of data fields, capturing contextual elements like dates, currency, percentages, profits, geographic countries, or rankings. The chart layer designates the layout type and maps these variables to specific visual elements like the horizontal and vertical axes, color codes, sizing markers, or facets.
Flint's built-in compiler evaluates these semantic inputs to automatically determine time-parsing rules, aggregation behaviors, numerical formatting, color palettes, and overall layouts. For instance, a financial profit variable automatically triggers a diverging color scheme to highlight positive versus negative shifts, while a calendar field seamlessly inherits the correct time-parsing and coordinate axis structures.
Alongside the core language, the developers launched a dedicated flint-chart-mcp server. The server allows artificial intelligence agents that support the Model Context Protocol to construct and preview data visualizations directly within an active dialogue, eliminating the need to leave the immediate chat interface.
In benchmark evaluations, researchers quoted in local Chinese media compared Flint against DirectVL, a traditional approach where artificial intelligence directly outputs an entire Vega-Lite configuration. Evaluated by large language models, Flint scored higher comprehensive marks across test batches using advanced models including GPT-5.1, GPT-5-mini, and GPT-4.1.