Flint: Microsoft Research's visualization language for AI

Microsoft Research introduced Flint, a visualization intermediate language that lets AI agents and people generate polished charts from compact, human-editable specifications. The open-source project ships with the flint-chart library and a flint-chart-mcp server, and a single Flint specification can compile to Vega-Lite, Apache ECharts, or Chart.js.

What is Flint and how does it work?

Flint is a lightweight specification language plus a compiler that derives low-level visualization settings from high-level intent. The compiler uses data specifications that capture semantic types and optional metadata, and a chart specification that maps fields to visual channels; from these the compiler derives parsing rules, scales, axes, aggregations, formatting, color schemes, layout, and a backend-native specification used to render the final visualization. Because the intermediate representation is separate from any single rendering library, the same compact spec can target multiple backends without rewriting the chart.

The system emphasizes semantic types to express data meaning, which helps the compiler choose appropriate scales, baselines, formatting, and color schemes. Flint also automatically manages sizing, spacing, labels, and layout so charts remain readable as cardinality and density change, removing the need for many explicit, fragile layout settings.

How does Flint help AI agents generate charts?

Flint reduces brittle, low-level specification work that tends to trip up LLMs and agents by exposing semantic types and a compact mapping of fields to channels instead of verbose, library-specific parameters. Field names, value patterns, and common data knowledge let agents infer whether a column is a date, price, percentage, or ranking; once meanings are declared, the compiler handles details such as temporal parsing, axis formatting, color scales, and cell sizing. The project also provides a Model Context Protocol server, flint-chart-mcp, so agents can create, validate, and render charts inside chat or coding environments with data embedded inline or read from configured local files.

How did Flint perform in evaluations?

In a research comparison against a baseline called DirectVL, which asks models to generate full Vega-Lite specifications, Flint received higher LLM-judge scores across three tested models on Tidy Tuesdays test data. Flint scored 16.27 versus 15.91 with GPT-5.1, 16.16 versus 15.60 with GPT-5-mini, and 15.91 versus 15.34 with GPT-4.1. The paper notes Flint's reliability enough that it now powers a Microsoft Research project called Data Formulator for AI-assisted data analysis and visualization.

Why it matters

Flint addresses a practical failure mode: agents and people can reliably express chart intent without hand-authoring many brittle, low-level parameters. That reduces the cognitive and engineering overhead of producing readable, trustworthy charts and makes agent-generated visualizations easier to inspect and edit. The multi-backend compilation approach also gives teams flexibility to render the same intent where a particular backend's features or integrations are preferred.

What to watch

Watch for contributions and forks on the GitHub repository (https://github.com/microsoft/flint-chart) and whether more tools adopt Flint's semantic-type layer. Also track its integration in Data Formulator and whether additional backends beyond Vega-Lite, Apache ECharts, and Chart.js are added to the compiler pipeline.

Project resources

Project site: https://microsoft.github.io/flint-chart/ GitHub: https://github.com/microsoft/flint-chart Flint MCP server instructions: https://microsoft.github.io/flint-chart/#/mcp

Authors credited on the project page include Chenglong Wang, Alper Sarikaya, Scott Tsukamaki, Michel Galley, and Jianfeng Gao.