Microsoft's open-source Flint turns one sentence into a full chart

Generating a chart from a sentence sounds simple until you try to make an AI do it. The problem isn’t the chart itself — it’s the avalanche of configuration options buried inside every visualization library. Tell a language model to output a complete Vega-Lite spec and you get back parameters that conflict, overlap, or just miss the point entirely.

Microsoft Research, in collaboration with Renmin University’s IDEAS Lab, has a cleaner answer. They’re releasing Flint, an open-source visualization intermediate language that separates what a chart is supposed to say from how a specific library renders it.

Flint splits its specification into two parts: data definition and chart definition. On the data side, you tag each field with its semantic type — year, month, price, percentage, profit, country, rank. On the chart side, you declare the chart type and map fields to visual elements like the X axis, Y axis, color, size, or facets. A built-in compiler takes it from there, deriving time parsing rules, aggregation strategies, number formats, color schemes, and layout without any manual poking.

The result is practical. A profit field automatically gets a diverging color scale that shows gains and losses. A date field picks up the right time parser and axis format. The compiler handles the grunt work that normally trips up LLMs when they try to produce raw library configuration.

Developers can also use the flint-chart-mcp server, which lets any MCP-compatible AI agent create and preview charts inside a chat window. No switching tools, no copying code.

In benchmarks against DirectVL — the baseline approach of generating complete Vega-Lite configurations — Flint scored higher across GPT-5.1, GPT-5-mini, and GPT-4.1. The gap was consistent enough to suggest the architecture itself is the advantage, not any specific model quirk.

Flint is already running inside Microsoft Research’s own Data Formulator tool. The company has open-sourced the flint-chart library, the MCP server, and documentation on GitHub under the microsoft/flint-chart repository.