LLM

LLM is a command-line tool and, at the same time, a Python library for working with large language models, created by Simon Willison. It talks to models from several vendors — the release post for version 0.26 names OpenAI, Anthropic and Gemini — and to local models served through Ollama, with support for many of them supplied by plugins such as llm-anthropic, llm-gemini and llm-ollama. Its author describes it as similar in shape to his other project, sqlite-utils, in being both a CLI tool and a library.

Capabilities

The tool is installed with ordinary Python tooling (the release post suggests pip, pipx or uv), stores API keys per vendor with llm keys set, has a configurable default model, and selects another model with -m.

Version 0.26 added tool support, which its author calls the biggest new feature since he started the project. In that version:

• --tool/-T exposes a named tool to the model and can be repeated; LLM ships simple demo tools such as llm_version and llm_time.
• Tool plugins add capabilities to whichever model is in use. The author’s first four are llm-tools-simpleeval (simple expression evaluation, for example mathematics), llm-tools-quickjs (a sandboxed QuickJS JavaScript interpreter whose state persists between calls), llm-tools-sqlite (read-only SQL against a local SQLite database) and llm-tools-datasette (SQL queries against a remote Datasette instance). The Datasette plugin is what he calls a “toolbox”: a plugin holding several tools, configured through a constructor.
• --functions takes a block of literal Python code and makes every function defined in it available to the model as a tool.
• --td (--tools-debug) prints each tool call and its response.
• In the Python library, model.chain() works like model.prompt() but executes the tool calls a response requests and prompts the model again with the results; an after_call argument lets the caller observe each call. Tools can be sync or async functions, and when a model requests several async tools at once they run concurrently.

Adoption & Ecosystem

The project is extended mainly through plugins, and at the time of 0.26 its author said he was most excited about the potential of tool plugins, keeping a cookiecutter template for writing them and documenting how model plugins add tool support. He also named acting as a Model Context Protocol client as clearly on the agenda, so that MCP servers could serve as additional sources of tools. He presents the tool-support release as making LLM a way to build “agents” in the sense of tools in a loop (see AI Agent and Function calling); the release itself is described in Large Language Models can run tools in your terminal with LLM 0.26.