This post announces version 0.26 of LLM, Simon Willison’s command-line tool and Python library for working with large language models, and the feature its author calls the biggest since he started the project: support for tools. With it, models from OpenAI, Anthropic and Gemini, and local models run through Ollama, can be given access to any tool that can be represented as a Python function, and new tools can be installed as plugins.
Most of the post is a walkthrough. It starts with trivial built-in demo tools, moves on to plugins that give a model a safe expression evaluator, a sandboxed JavaScript interpreter or SQL access to a local SQLite database or a remote Datasette instance, then shows ad-hoc tools defined by passing literal Python code on the command line, and finally the Python API. Along the way it shows models recovering from their own mistakes through tool feedback — retrying a calculation after a function turned out not to exist, and fetching a database schema after a guessed SQL query failed.
The closing sections step back. The author explains why tool support took him so long, gives his view of the term “agents”, and lists what comes next, including acting as a Model Context Protocol client.
Key Points
- LLM 0.26 lets the CLI and the Python library grant models access to tools written as Python functions, across OpenAI, Anthropic, Gemini and Ollama-served local models.
- Tools can come from plugins, loaded by name with
--tool/-T; a “toolbox” is the author’s name for a plugin that holds several tools and is configured through a constructor, such as a Datasette toolbox pointed at a particular instance. - The
--functionsoption accepts a block of Python code and exposes every function it defines to the model as a tool — which the author calls “such a hack”, demonstrating it with a four-line blog-search tool that returns raw HTML. - The Python API gains
model.chain(), which spots tool call requests in a response, executes them and prompts the model again with the results, potentially across many responses. - The author argues, from his own experience of tracking the area, that tool use has become the single most effective way to extend the abilities of language models, and that the underlying trick is simple: tell the model which tools exist, let it output special syntax requesting one and stop, run the tool, and prompt again with the result.
- He says he held off because building an abstraction that works across many models needed vendors to agree on how tool use should work, and that there is now a very definite consensus among them.
- On terminology, he still does not like the word “agents”, but observes that the field appears to be converging on “tools in a loop” — which, he says, is what this release provides.
Context
The post is written by the tool’s own author and reports his own demonstrations, so its claims about what works are one developer’s hands-on account rather than a measured comparison. It treats tool use as the same pattern whether a vendor calls it tool usage or Function calling, and it frames the release as the step that makes LLM usable for building what others call agents (AI Agent). The author links his earlier interest in the idea to the ReAct pattern, of which he had once built a small implementation himself.