Skip to content
Main Site News Console

Datasette Agent

· Simon Willison Translated
技巧LLM

We just released the first version of Datasette Agent, a new, extensible AI assistant for Datasette. Over the past three-plus years I’ve been building my LLM Python library, and Datasette Agent is where LLM and Datasette finally come together. I’m incredibly excited about it!

Datasette Agent provides a conversational interface that lets you ask questions of data stored in Datasette. With the datasette-agent-charts plugin installed, it can also generate charts for your data.

The demo

The announcement post (published on the new Datasette project blog) includes this demo video:

The video was recorded against the new agent.datasette.io hosted demo instance, which runs on Datasette Agent and is connected to a handful of example databases, including the classic global-power-plants dataset from WRI, as well as a copy of the Datasette backup of my blog.

This hosted demo runs on Gemini 3.1 Flash-Lite — it’s cheap, fast, and has no trouble writing SQLite queries.

The question I asked in the demo was:

When did Simon last see a pelican?

It ran this SQL query:

SELECT
  title, commentary, created
FROM
  blog_beat
WHERE
  beat_type = 'sighting'
  AND (title LIKE '%pelican%' OR commentary LIKE '%pelican%')
ORDER BY
  created DESC
LIMIT
  5

And then it answered:

Simon last recorded seeing a pelican on May 20, 2026.

The sighting included a California brown pelican, along with a common loon, a Canada goose, a striped shore crab, and a California sea lion.

That’s the entry on my blog, along with a Markdown export of the full conversation.

The plugins

One of my favorite things about Datasette Agent is that, like the rest of Datasette, it can be extended with plugins.

We’ve already released three plugins:

It’s been really fun developing plugins. I already have a bunch of prototypes, though none of them are quite at alpha quality yet.

Claude Code and OpenAI Codex have both been excellent at writing plugins — just point them at a checked-out copy of the datasette-agent repository as reference material and tell them what you want to build!

Running it against local models

I’ve also been having a lot of fun using the new plugins against local models. Here’s a one-line uv command that runs the plugin against gemma-4-26b-a4b via LM Studio on a Mac:

uvx --prerelease=allow \
  --with datasette-agent --with llm-lmstudio \
  datasette --internal internal.db --root \
  -s plugins.datasette-llm.default_model lmstudio/google/gemma-4-26b-a4b \
  data.db

Datasette Agent needs reliable tool-calling capabilities, as well as the model’s ability to generate SQL queries that work against SQLite. The open-weight models released over the past six months are getting increasingly good at both of those things.

What’s next

Datasette Agent opens up countless opportunities for LLMs and for the Datasette ecosystem as a whole.

It has already influenced the major refactor in LLM 0.32a0; I’m also just about ready to push that into a stable release, likely with some additional “LLM agent” abstractions extracted from Datasette Agent itself.

I’ve been exploring my own implementation of Claude Artifacts, and it’s starting to take shape as a plugin.

I’m excited to build my own Claw with Datasette Agent — a personal AI assistant built around data imported from different parts of my digital life; it also gives me a great excuse to revisit my earlier Dogsheep family of tools.

We’ll also be rolling out Datasette Agent to Datasette Cloud users.

If you’d like to discuss the project, join us in the #datasette-agent Discord channel.