If anyone has earned the right to combine AI with databases, it’s Simon Willison. He’s been working on his LLM Python library for over three years, and on Datasette, his SQLite explorer, for even longer. Now both projects finally come together: Datasette Agent.
What Datasette Agent does
At its core: you ask questions in natural language, and the agent writes SQL queries, executes them, and delivers answers. Willison’s demo example: he asks his blog backup when he last saw a pelican — and gets a complete answer with date and details.
That alone would be nice but not revolutionary. What makes Datasette Agent special: it’s extensible through plugins.
The plugin system
Three plugins ship at launch:
- datasette-agent-charts: Generates charts with Observable Plot
- datasette-agent-openai-imagegen: Creates images from data using ChatGPT Images 2.0
- datasette-agent-sprites: Executes code in a persistent Fly Sprites sandbox
Cleverly designed. Willison notes that Claude Code and OpenAI Codex are both excellent at writing new plugins — just point them at a checkout of the datasette-agent repo for reference and tell them what to build.
Local models work
Particularly interesting: Datasette Agent runs with local models too. Willison shows a one-liner using uv that runs Google’s gemma-4-26b through LM Studio. The open-weight models from the last six months can reliably handle tool calls and write SQLite queries — something unthinkable a year ago.
What’s next
Willison plans integration into Datasette Cloud and is building his own personal AI assistant called ‘Claw’ — built on imported data from different parts of his digital life. This ties back into his older ‘Dogsheep’ project, which imports data from various sources into SQLite.
A live demo is already running at agent.datasette.io — powered by Gemini 3.1 Flash-Lite, chosen because it’s cheap, fast, and handles SQLite queries without issues.
Sources: Simon Willison — Datasette Agent