Flint — Microsoft's Open-Source Language That Compiles Visualizations for AI Agents

Microsoft Research released Flint — an open-source language for data visualization specification designed for the era of AI agents and copilot applications. The project addresses a long-standing engineering challenge: how to enable AI agents to produce precise, visually polished charts without verbose, error-prone specification files.

The Tradeoff Flint Resolves

Modern visualization libraries — Vega-Lite, Apache ECharts, and Chart.js — offer a high degree of customization, but at the cost of verbose specification files. When an AI agent must generate such a detailed chart description, even a minor syntax or logic error can result in a non-functional or visually poor output. On the other hand, short specifications that rely on system defaults rarely produce visualizations suitable for business or research use.

Flint is a concise middle ground: a compact specification readable enough that a language model can reliably produce an accurate record, yet rich enough in semantic information that the compiler can automatically derive everything missing — scales, axes, aggregations, formatting, and color schemes.

How Does the Compiler Work?

The Flint compiler takes two inputs: a data specification describing semantic types and metadata, and a chart specification defining the chart type and field-to-visual-channel mapping. From these inputs, the compiler automatically derives parsing rules, scales, axes, aggregations, formatting, color schemes, and element layout — then generates a native specification for the chosen backend.

A single Flint specification can be compiled simultaneously into Vega-Lite, Apache ECharts, or Chart.js, giving organizations the flexibility to choose a rendering backend according to their own needs without writing separate specifications.

Semantic Types as the Core Mechanism

The central concept of Flint is semantic data types. Rather than the user or agent manually defining every aspect of the display, Flint recognizes types such as dates, percentages, prices, or rankings and intelligently infers the appropriate scale, axis label format, and color scheme from that information. When data changes in cardinality, Flint automatically adjusts element layout and sizes.

This means that an agent generating a Flint specification does not need to “know” how Vega-Lite formats a time axis or how ECharts handles negative values — Flint takes on that responsibility.

Testing Results on Language Models

Microsoft Research tested Flint on three language models. In all cases, charts generated via Flint received higher evaluation scores than charts generated using the direct Vega-Lite approach (DirectVL). The results confirm that a more concise and semantically richer format reduces the error rate in AI-assisted visualization — a key assumption for scalable copilot applications with data display components.

MCP Integration and Availability

Flint ships with an MCP server that enables chart creation directly within chat interfaces supporting the Model Context Protocol. The project is integrated with the Data Formulator tool for AI-assisted data analysis.

Code is available at github.com/microsoft/flint-chart, and the project page at microsoft.github.io/flint-chart/. The authors are Microsoft Research researchers: Chenglong Wang, Alper Sarikaya, Scott Tsukamaki, Michel Galley, and Jianfeng Gao.

Summary

Flint does not replace Vega-Lite or ECharts — it compiles into them. Its contribution lies in the abstraction layer that makes visualization specifications simultaneously accessible to generative models and editable by humans. For teams building AI copilot applications with data visualization components, Flint offers a concrete engineering solution to a problem that previously often required manual mediation or a compromise between quality and reliability.