Datasette Agent gets chat panel to build database apps with prompts

datasette-agent 0.5a0, released by Simon Willison on November 3, adds a chat panel that sits next to Datasette tables and apps, making it possible to build and adjust databases using only prompts.

What's new in this release

Simon Willison published version 0.5a0 of datasette-agent on November 3, 2026, an LLM-powered agent that runs on top of Datasette, his open-source project for exploring and publishing data in SQLite. The a0 suffix follows Python's versioning convention (PEP 440): it's an explicit alpha, even before a stable 0.5, a sign that the project is still under heavy construction.

The central feature of the release is something called "the panel": the agent's chat is no longer an isolated screen and instead opens next to Datasette's own pages, such as a table or the result of a SQL query. In practice, the developer no longer leaves the data context to talk to the AI.

Chat next to the page, not in another tab

The point Willison highlights in the announcement is exactly this juxtaposition between the data interface and the chat interface:

it lets the agent chat open and then sit next to pages from Datasette, for example a table or query page.

Simon Willison

This design choice matters more than it seems. AI tools for data usually work as a separate chat that generates a query and returns the result in another window, breaking the flow of someone investigating a table. Placing the agent next to the page itself eliminates that back-and-forth: the question, the answer, and the data stay in the same field of view.

From prompt to working app

The second announcement in the release is the combination with datasette-apps, another Willison project integrated into Datasette. According to Willison, this combination makes it possible to "prompt" a database-backed app into existing, and then keep modifying that app through the same chat sitting next to it.

This describes a vibe coding cycle applied to data: instead of writing the schema, the query, and the screen by hand, the developer describes what they want in natural language, watches the app come to life, and requests incremental adjustments without switching tools. It's the same pattern Willison has been exploring on other recent fronts, such as the post about the inflection point of game development through vibe coding, published on October 28, and the experiment with the Qwen 3.8 27B model running locally, from that same November 3. datasette-agent applies this logic to the domain of databases and internal dashboards.

Why this matters to people who already work with Datasette

People who use Datasette are usually solving a specific problem: exposing a SQLite database for querying, building a quick read-only API for a public dataset, or giving a non-technical team a way to explore a table without writing SQL. The agent attacks exactly that frontier between "I have the data" and "I need an interface over it".

In short: the panel reduces the distance between exploring a table and building something on top of it. Before, that required opening an editor, writing a plugin, or putting together a separate page; now the request can come straight from the chat next to the data. For internal prototypes, single-use tools, and exploring a new dataset, this cuts real steps from the workflow.

What still weighs against using this in production

Willison's announcement is direct and short, with no benchmark, no detail about which model runs behind the agent, and no data policy published in the post itself. That leaves open questions any team needs to answer before connecting this to a real database:

• Security model: an agent with the power to create and modify an app on top of a database is, by definition, an agent with write access to application logic; the published material gives no detail about limits, review, or sandboxing of this process.
• Maturity: the a0 suffix is flagged as alpha by the versioning scheme itself; breaking changes between this version and a future stable 0.5 are expected.
• Dependence on a personal project: both Datasette and datasette-agent are maintained mainly by Willison; this is a common pattern in the Python open-data ecosystem, but it's a continuity risk factor corporate teams tend to weigh.
• Account-based access: trying it requires logging in with a GitHub account at agent.datasette.io, which means exposing that identity to the hosted service, and not necessarily running it locally.

How to try it

The way to try the feature, according to the announcement itself, is to log in with a GitHub account at agent.datasette.io. Willison didn't detail usage limits, pricing, or additional requirements in the launch post, so it's worth treating the test for what it is: an alpha preview of a project that's still defining its own final shape.

For those who already have Datasette instances running, with internal or public data, the reasonable path is to set aside a low-risk dataset for this first contact: the panel's gain shows up quickly in exploration, but the leap to "letting the agent write the app" deserves the same skepticism any LLM code generation calls for, human review before going to production.