A friendly robot analyzes glowing line and bar charts on a futuristic data dashboard.
Microsoft Research has published Flint-interactive, an interaction layer for Flint, its open-source chart language for AI agents. It aims to fix a common problem. An agent can draw a chart, but it often struggles to make that chart respond reliably to clicks, brushes and filters. The post is dated October 22, 2026. Its authors are Chenglong Wang, Yuwei Xiao, Alper Sarikaya, Michel Galley, Chris Quirk and Jianfeng Gao. Sarikaya is listed as a Power BI engineer.

A friendly robot analyzes glowing line and bar charts on a futuristic data dashboard. Where Flint-interactive fits​

Flint is not new. Microsoft describes it as an open-source visualization language that lets AI agents create expressive charts from compact, human-editable specifications. One Flint specification can compile to Vega-Lite, Apache ECharts or Chart.js without rewriting the chart. The project ships the flint-chart library and the flint-chart-mcp server, so agents can create, validate and render charts in chat or coding environments.

Flint-interactive extends that idea to behavior. Microsoft's pitch is that static charts are fine for a first look, but readers want to zoom, compare views and have a selection in a chart highlight related text or table rows. Microsoft says agents rarely get a satisfying interactive chart in one turn. Repeated rounds of reasoning, coding and debugging cost tokens and time. Fixing one layout or interaction bug can break something else. Those are the authors' stated rationale, not independently measured failure rates.

How the model works​

Flint Canvas is described as a state machine. It tracks user interactions and manages rendering state. The design splits the work in two:

  • Visual logic, handled by the compiler. Raw triggers such as clicks, drags and key presses become semantic events that identify chart elements and data. The compiler also does hit testing and turns update operators into canvas changes.
  • Interaction logic, written by the user or the agent. The author maps semantic events to update operators. Examples include set style to highlight data, set annotation to label a point, and set viewport to pan or zoom. The same events can drive external tables, widgets or application state.

The practical difference is that you describe what was selected and what should change. You don't reach into renderer-specific marks. Microsoft argues this makes interactions easier to write and maintain. The post gives no benchmark data for that claim, so treat it as a design argument.

Three ways to use it​

1. In chat. Flint-MCP and MCP Apps let you use it in a chat or canvas environment. According to Microsoft, the agent can produce an interaction_spec alongside the Flint chart spec and render the interactive chart in the conversation. A selection can be passed as context to a follow-up question.

2. In your own app. Flint Interactive Canvas emits semantic events. Your event handlers can filter a table, update another chart or change application state. In the other direction, your widgets can send update operators to the canvas. Microsoft's example is a table and chart that highlight each other.

3. Bespoke interactions. You define a trigger and a handler function through the programming API. Microsoft's example highlights a clicked country and labels it by combining set-style and set-annotation. Its most elaborate demo is a DimpVis-style example. Dragging a point along a time path shows how cities' temperature and precipitation change.

The post also shows a preset called linked-brush. A selection in one view highlights matching groups in another.

What the repository adds​

The Microsoft Flint repository on GitHub says Flint supports 52 chart types across five rendering backends: Vega-Lite, ECharts, Chart.js, Plotly and native Excel charts. It lists 14 themes and 23 interaction presets. The repository describes two ways to use it. One is an MCP server for everyday agent use. The other is an npm library for building interactive apps. These figures come from the repository, not the October 22 blog post, and may change.

The library documentation draws a line that matters to developers. Interactive rendering is opt-in and ships separately from the static assembly entry point. The documented API uses mountChart from flint-chart/interactive, with a React wrapper under flint-chart/react.

The documentation also notes backend limits:

  • For Vega-Lite, interactions are enabled explicitly through an interactions option or the spec's interaction_spec. With neither, the chart is static.
  • Other backends render the chart static and drop its interactions with an interactions_ignored warning.
  • Static SVG and PNG rendering and Excel output don't run the interactive surface.

In practice, an existing static Flint workflow won't become interactive by itself. You have to wire interactions in. The blog's examples also shouldn't be read as proof that every interaction works on every chart type or renderer.

Context and caveats​

Flint's first public appearance came earlier this year. Better Stack's guide dates the project to July 2026 and calls it early in development. It says the chart coverage at that time didn't include maps, 3D charts, network graphs or complex layering. Accessibility was on the roadmap but not yet implemented. That guide predates Flint-interactive, so check the current repository before assuming those gaps remain.

Some outside coverage is looser than the evidence. One site suggested Flint could eventually influence chart creation in Excel and Power BI. Microsoft's post doesn't say that. The verifiable links are that a Power BI engineer co-authored it and that native Excel charts are one of the five backends.

Who should care​

  • Developers building agents or analytics apps get a documented way to separate interaction logic from rendering code. They also get an open-source path to try it.
  • Power users working in chat tools with MCP support can ask for linked, selectable charts and feed selections back into the conversation.
  • Skeptics should wait for independent testing of the reliability and cost-saving claims. Microsoft's post offers none.

The release is real, open source and documented. The benefits are Microsoft's claims until others test them.

 

References

  1. Flint-interactive: A language for creating interactive data visualizations - Microsoft Microsoft 2026-10-22T16:00:00+00:00
  2. Flint: A Visualization Language for the AI Era github.com
  3. README.md github.com