Datasette Agent: Conversing with Your SQLite Data

For three years, Simon Willison has been developing LLM, his Python library for working with language models from the command line. On May 21, 2026, he announced on his blog the first release of Datasette Agent, the point where LLM and Datasette finally converge into a single product. This is no minor integration: it's the piece that turns Datasette, the most widely used SQLite data explorer in the Python world, into something you can have a conversation with.

The official announcement on the Datasette blog includes a demo video showing the basic flow: a user types a question in natural language, the agent translates it to SQL, executes the query against the Datasette instance, and returns a readable answer. No SQL required.

What is Datasette Agent and how does it work

Datasette is an open-source tool for exploring, publishing, and sharing SQLite databases through a web interface. It has a solid user base among data journalists, researchers, and small data teams who need to publish datasets without complex infrastructure.

Datasette Agent adds a conversational layer on top. Internally, it relies on Willison's LLM library, which acts as an abstraction layer over different language models. This means the agent isn't locked to a specific provider: whoever configures the instance decides which model powers it underneath.

Beyond the conversational interface, the project introduces its first official plugin: datasette-agent-charts. When installed, the agent can generate data visualizations directly within the same conversation. Questions like "how have sales evolved by quarter?" return a chart, not just a table.

Why the extensible architecture matters

From a technical perspective, what's most relevant isn't the chat itself, but that the system is designed to be extensible from the ground up. Datasette has had an active plugin ecosystem for years, and Datasette Agent inherits that philosophy: any additional functionality can arrive as a plugin without touching the core.

This matters because Datasette's use cases are highly diverse. A municipal budget dataset, a file of anonymized medical records, or a product catalog have completely different structures and visualization needs. A rigid architecture would fail in the face of that diversity; an extensible one lets the community solve their own problems.

It's also worth noting that LLM, the library powering the engine, has three years of iterative development behind it and supports multiple backends. That gives Datasette Agent a more mature foundation than the version number (first release) might suggest.

Who finds this useful today

Datasette Agent is especially interesting for profiles working with data who don't master SQL: journalists, public policy analysts, NGO staff managing their own datasets. Also for technical teams who want to offer exploratory access to internal data without building a custom interface.

That said, we're looking at a first version. The promised extensibility will depend on how many plugins the community develops in the coming months. Datasette's plugin ecosystem is active, but building agent plugins—with tools, context, and error handling—requires more care than a conventional visualization plugin.

From ClaudeWave, we're watching with interest how established data tools begin incorporating conversational layers without abandoning their original philosophy. That Willison waited three years to bring LLM and Datasette together, when he could have done it sooner, suggests he waited until the foundation was solid enough. In this sector, that's usually a good sign.