datasette-agent 0.6a0 adds an agent prompt to the Datasette homepage
The 0.6a0 alpha release of Datasette Agent adds a prompt box to the Datasette homepage and supports hash-mentions of tables and queries.
Datasette Agent, the LLM-powered plugin that lets people ask questions about their data in Datasette, now puts a chat box on the Datasette homepage. Simon Willison announced the change in datasette-agent 0.6a0, a pre-release dated 15 November 2026. He described the rest of the release as accumulated small fixes. The new section sits on the homepage under the heading "Datasette Agent" and appears only for users who hold the datasette-agent permission. It lets a user start a new conversation without first opening a table or a saved query.
The release notes say that typing # in the homepage chat box tags tables, views, databases and stored queries. This hash-mention feature works the same way in the agent chat box and matches the behaviour on the /-/agent page. The notes don't say how the model resolves a tag internally.
The 0.6a0 notes also cover several changes to the command line and chat behaviour. The datasette agent chat command now accepts -c or --config to load a Datasette JSON or YAML configuration file. It also takes a repeatable -s or --setting option for single settings, and both work as they do in datasette serve. The -m option overrides a default_model set in either place. Without a configured default, background agents, the command-line chat and conversations created through the API now use the first available model, which is the one the model picker preselects. A default that is configured but cannot be used still produces an error.
Two fixes concern reliability. The notes say every user message was sent to the model twice, in both chats and in background agents, and that bug is now fixed. A conversation whose model has since been blocked, removed from the allowlist, or lost its API key used to keep running on that model. It now stops with an error asking the user to start a new chat. The "View SQL query" link under a result now passes the query's named parameters.
Other changes deal with links and the panel. The system prompt now tells the model to use root-relative paths when linking to pages on the same instance, because the notes say models were inventing hostnames such as https://datasette.local/. Links from pages in the panel to other sites now ask for confirmation and open in a new tab. Forms in the panel that submit to another site are blocked. A bug that made http:// links do nothing when Datasette was served over https, for example behind a TLS-terminating proxy without the force_https_urls setting, is fixed under issue #54. Pages shown in the panel no longer display the "Powered by Datasette" footer.
The release page doesn't say which models the agent supports. It also doesn't describe how the homepage prompt behaves for users without the datasette-agent permission beyond not showing it to them.
What changed
datasette-agent 0.6a0 adds an agent chat box to the Datasette homepage for users with the datasette-agent permission, and supports # hash-mentions of tables, views, databases and stored queries in it.
Type: New capability · Takes effect: Sunday, November 15, 2026 · Applies to: datasette-agent 0.6a0
Why it matters
Teams that use Datasette for internal data exploration get a more direct way to put an AI agent in front of their data. Because this is an alpha, its behaviour may change before a stable release.
What you can do
Teams interested in agent-driven data exploration can test the alpha on a non-production Datasette instance before relying on it.
Original source
Where we saw it
Simon Willison
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works