Flint Agent
GitHub
Microsoft released Flint, a compact chart-spec language that lets AI agents generate polished visualizations without fiddling with axes and layout code. It's an MCP server too, so agents can build and render charts mid-conversation across Vega-Lite, ECharts, Chart.js, Plotly, and even native Excel.
Based on reporting by GitHub — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Charting libraries have always made a strange demand of AI agents: describe a bar chart in exhaustive detail, down to tick spacing and legend placement, or get something ugly and wrong. Microsoft Research's new project, Flint, built with Renmin University's IDEAS Lab, tries to remove that burden entirely. Instead of hand-tuning scales and margins, you hand Flint a short spec describing what the data means, and the compiler figures out the rest.
The trick is semantic typing. Flint recognizes more than 70 field categories, things like Rank, Temperature, Price, or Country, and uses that context along with chart type and data cardinality to automatically pick sensible layouts, spacing, and labels. A scatter plot comparing car weight to mpg, colored by country of origin, becomes a dozen lines of JSON rather than a sprawling configuration object. That compactness matters for agents especially, since large language models are far more reliable at producing short, structured specs than verbose ones full of numeric knobs.
What makes Flint more than a syntax trick is that one input compiles to five different native outputs: Vega-Lite, ECharts, Chart.js, Plotly, and, notably, actual Excel charts via Office.js. Swapping backends is a one-line function change, not a rewrite. As of version 0.4.0, released July 24, 2026, Flint added 38 Plotly chart types and 18 editable native Excel templates, and a few days earlier it introduced dynamic widgets that let users switch chart types and tweak properties after the fact.
The MCP server, flint-chart-mcp, is arguably the more interesting half for anyone building agent workflows right now. It lets an agent create, validate, and render a chart inside the same chat session where the question was asked, either as an interactive view or a static PNG. Data can come from inline rows or straight from local CSV, JSON, or TSV files, which sidesteps a lot of the copy-paste friction agent tooling usually creates.
Flint is MIT-licensed and lives on GitHub under Microsoft's org, with a JavaScript/TypeScript package shipping now and a Python port still source-only. A research paper is reportedly on the way, but even without it, this reads like a quiet bet that the future of agent-generated visualization isn't smarter prompting, it's a better intermediate format.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous kind of AI infrastructure that actually matters more than another chatbot demo: a format designed so models fail less often, not one designed to look impressive in a keynote. I'd rather see ten of these boring compiler-shaped tools than another benchmark-chasing model release, and the multi-backend approach, including plain Excel, shows someone actually thought about real workplaces instead of just developer Twitter.
Read more about this at: GitHub
