source: simon willison: datasette agent
level: technical
datasette agent is a new ai assistant for datasette that provides a conversational interface for asking questions about stored data. it integrates the llm python library with datasette, allowing users to type natural language queries that get translated into sql. the tool can run against local models or cloud apis, and it works with sqlite databases. a live demo uses gemini 3.1 flash-lite to answer questions from example databases, including a blog backup and global power plant data.
the assistant is extensible through plugins. three initial plugins add chart generation with observable plot, image generation via chatgpt images 2.0, and code execution in a fly sprites sandbox. building plugins is straightforward, and tools like claude code and openai codex can help write them. the system requires models that reliably produce sql queries and tool calls, and recent open weight models are increasingly capable of handling these tasks.
datasette agent opens new possibilities for the llm and datasette ecosystem. it has already influenced a major refactor of the llm library, and the developer is exploring a claude artifacts-like plugin. future plans include using it to build a personal ai assistant that integrates data from various digital sources, and rolling it out to datasette cloud users. a discord channel is available for community discussion.
why it matters: it makes querying structured data as simple as asking a question, lowering the barrier for non-technical users to explore databases and enabling rapid prototyping of ai-powered data tools.
source: simon willison: datasette agent