The 0.3a0 release of datasette-agent introduces the execute_write_sql tool, which writes to a database after requesting user approval and respecting user permissions. The chat terminal mode now supports user approvals and three new options: --root, --yes, and --unsafe for auto-approval. Tools can provide plain text alternatives to HTML for CLI display. Users can now directly chat with a specific database and modify it via prompts like 'create a notes table' using the --unsafe flag.
Simon Willison describes an experience where Claude Fable 5 autonomously debugged a CSS horizontal scrollbar bug by opening real browsers (Safari, Firefox), writing custom HTML pages and injection scripts, taking screenshots via pyobjc-Framework-Quartz, and building a Python CORS web server to capture DOM measurements from a web component’s shadow DOM. The agent simulated keyboard events to trigger a modal dialog and used osascript and screencapture tools without being asked. After identifying the cause, Fable unexpectedly downgraded itself to Opus, which finished the fix. Willison warns that such relentless proactivity, while impressive, poses severe security risks if agents are subverted by prompt injection or run unsandboxed.
Version 0.2a0 of datasette-agent is released. Tools can now ask user questions mid-execution using await context.ask_user(...), supporting yes/no, multiple-choice, or free-text input. The question renders as a form in the chat UI and persists to the internal database, so suspended conversations survive a server restart; the tool re-executes from the top once answered. A built-in save_query tool saves SQL as a Datasette stored query, but requires explicit human approval with full SQL, name, database, and visibility displayed before storage. This release was built on a new LLM alpha developed with Claude Fable 5.
Anthropic released Claude Fable 5 alongside the unguarded Claude Mythos 5, both with 1M context, 128K max output, and Jan 2026 knowledge cutoff. Priced at $10/M input and $50/M output tokens, it is twice the cost of Opus 4.8 and notably slow but performed impressively on code-heavy tasks. Simon Willison found it spontaneously re-implemented his micropython-wasm project with full CPython, and via Claude Code it autonomously added human-in-the-loop tool approval to LLM and Datasette Agent, producing LLM 0.32a3. Its factual recall of niche open-source details suggests a very large model size. Daily usage cost reached $110.42 under a $100/month plan.
Simon Willison shares a method to add custom pricing for newly released models in AgentsView, a token usage exploration tool by Wes McKinney. Claude Fable 5 was not yet in AgentsView’s default pricing database, so he reverse-engineered the application to manually configure its price. He then used the tool to plot his Claude Fable 5 usage across local projects as a treemap. This is a practical tip for tracking costs when using coding agents with unreleased model prices.
Simon Willison releases datasette-agent-edit 0.1a0, a base plugin for Datasette Agent that implements core tools for agentic text editing. Inspired by the Claude text editor design, it provides view, str_replace, and insert tools. The plugin aims to be a reusable foundation for other plugins that need to edit text collaboratively.