FlexViz: open-source agent-native data viz where agents read your dashboard as structured state, 1B points in 251ms

Adorable-Giraffe5754 · reddit · 2026-09-22

A developer released FlexViz, a free open-source Python library for interactive exploration of large datasets that is explicitly agent-native: the agent and the human work on the same dashboard, communicating via structured state rather than screenshots.

How it works: ask the agent to explore a Parquet file and it inspects the data, builds a dashboard, and hands you a link. When you zoom, brush, or cross-filter, the agent can read that exact view back as structured state and analyze the actual rows behind your selection. It also tracks the views you discussed, so you can revisit previous explorations and have it draft a report of the analysis.

Performance: an optimized in-memory query engine plus polars-based out-of-core support keeps it interactive beyond RAM — a 1-billion-datapoint (5×200M) line chart renders in 251 ms in-memory and 4.2 s straight from Parquet.

Code and demo are public; the author is asking whether the human↔agent loop fits better as a skill or an MCP server.

Original post →

More from coding & agent

coding & agent channel →