Simon Willison has shipped datasette-llm 0.1a9, a small update to the plugin that connects Datasette to large language models. The change is simple: the plugin now uses whatever default model you’ve already set in LLM, Willison’s command-line tool for working with language models. Willison says it’s “a tiny user experience improvement” while he works on “big upgrades to Datasette Agent” that are still to come.
It’s a small release, but it makes Datasette Agent quicker to try. Here’s what changed and why it matters.
One less flag to remember
Before this release, starting Datasette Agent meant telling it which model to use. You had to add a setting like -s plugins.datasette-agent.default_model gpt-6-astra to your command.
Now you don’t. Willison says this one line is enough:
uvx --with datasette-agent datasette content.db --internal internal.db
The plugin checks LLM’s configuration and uses the model you’ve already picked there.
Your LLM settings carry over
This works because datasette-llm sits on top of LLM. LLM lets you set a default model once, and every tool built on it can then use that choice. If you’ve already set up LLM with your preferred provider and API keys, Datasette Agent now inherits those settings automatically.
Willison’s tools all fit together this way. LLM handles model access, Datasette handles data, and plugins connect the two. When one layer gets smarter about its defaults, the layers above it get easier to use.
Trying it costs almost nothing
The command uses uvx, from the uv Python toolchain. It runs a tool in a temporary environment without a permanent install. Add the new default-model behavior and you can go from nothing to an AI agent working on a SQLite database (content.db in the example) with one command.
The --internal internal.db part points Datasette at a separate database it uses for its own internal state. Your content database stays separate.
Datasette Agent is the real story
What stands out here is the hint about what’s next. Willison says he has “big upgrades to Datasette Agent in the pipeline,” and this release clears the way for them. Making setup easier first is a smart move. Any new features will reach more people if trying them out doesn’t start with a configuration error.
Datasette Agent fits a wider trend: people want to ask questions about their data in plain English instead of writing SQL. Many commercial tools offer this as a hosted service. Willison’s version is open source, runs locally, and works with whichever model provider you already use through LLM.
It’s still alpha software
The version number makes this clear. 0.1a9 is the ninth alpha of a 0.1 release, so expect rough edges and changes ahead. This is software for experimenting, not for running production workflows yet.
It also assumes you’ve already set a default model in LLM. If you haven’t, you’ll still need to configure one somewhere. The change saves you a step, but you still need working LLM credentials first.
What comes next
This is a quality-of-life release, and it’s useful for that reason alone. Developers often drop a new tool after one failed setup attempt, so removing a required flag matters more than it looks. The more interesting news will come with the Datasette Agent upgrades Willison has promised. Details on this release are available in Simon Willison’s original post.