Datasette Releases LLM-Powered Agent Plugin for SQLite

Datasette Agent is an open-source plugin for Datasette released on 21 May 2026, according to the Datasette project blog. The project blog describes the plugin as an extensible AI assistant that integrates the datasette-llm adapter into Datasette, enabling tool-calling model access across frontier vendors and open-weight models. The GitHub repository documents the interactive routes where the assistant is available, for example /-/agent, plus background-agent endpoints under /-/agent/background and /-/agent/explore/. The README lists three separate permissions: datasette-agent, datasette-agent-explore, and datasette-agent-background, and describes how saved SQL queries created by the agent require explicit human approval and run through Datasette's existing query analysis and store endpoints.

Per the project blog and README, Datasette Agent uses datasette-llm as its model-calling layer. The documentation includes a quickstart example using gpt-5.5 as the default model (Datasette blog), with support for local models such as lmstudio/qwen3.5-9b and OpenAI Codex via openai-codex/gpt-5.5 (Datasette blog). The plugin supports three user-facing flows: an in-browser conversational chat (/-/agent), an "Explore with AI agent" that inspects a selected database or table and writes a report, and background agents that are spawned with a goal and run autonomously until stopped. The README explains the save_query tool and the permission checks that gate SQL execution and persistence, referencing Datasette's existing /-/queries/analyze and /-/queries/store endpoints for validation and storage.

Companies and teams combining data cataloging or lightweight database UIs with LLM-driven querying are increasingly common. Datasette Agent follows that pattern by embedding an LLM-driven conversational layer directly into a database browser. For practitioners, this reduces friction when prototyping natural-language interfaces over SQLite datasets and provides a structured path for LLMs to propose and persist SQL, while preserving human approval steps for query storage. The plugin's support for both hosted vendor models and local open-weight models reflects a broader tooling trend that prioritizes interchangeability of model backends.

For adopters and observers, watch for:

• •integration patterns with existing Datasette plugins such as visualization extensions
• •how teams configure model backends for cost, latency and data governance
• •operational details for running background agents in production-like environments. The Datasette blog hosts a live demo (requires GitHub sign-in) and a video walkthrough that demonstrate typical queries and chart export features, per Simon Willison's newsletter and the project blog

Editorial analysis: From a tooling perspective, the plugin's explicit permission model and the save_query human-approval flow are important for teams that need auditability when an LLM generates SQL. Industry-pattern observations: Projects exposing LLMs to execute SQL commonly implement approval or dry-run steps to avoid unintended writes; Datasette Agent does this by routing persistence through Datasette's own endpoints and permission checks. Implementation details for deployers are in the GitHub README and the project blog, which include quickstart examples and configuration snippets.