A Free Command Line Tool Now Shows You What a Model Is Thinking, and Talks to Almost Any Model

Simon Willison has released version 0.32 of LLM, the free open source tool he has been building for years, and he calls it the most significant update since the project launched. LLM is a command line utility: you type a prompt into a terminal, it sends it to whichever model you have configured, and it logs everything to a local database. Think of it as a plain, scriptable version of a chat window that happens to work with every major provider at once.

The headline feature is that reasoning models now show their reasoning traces. Reasoning models are the ones that work through a problem step by step before answering, and until now those steps were hidden. LLM prints them to standard error, which is the technical way of saying they appear on your screen but do not end up in the output if you pipe the answer into another program. If you would rather not see them, the flag is -R. The default model for a bare prompt is now GPT-5.6 Luna, chosen because it is cheap and capable rather than because it is the biggest.

Two other changes matter more than they sound. The first is server-side tools: the model provider runs the tool on their own infrastructure, so a single command can now ask OpenAI to execute Python in its code interpreter or run a web search, and the companion llm-anthropic plugin adds web search, web fetch, code execution and an MCP connector. The second is a new llm openai endpoint command that fires a prompt at any OpenAI-compatible address as a one-liner. That includes a model running locally on your own machine through LM Studio, which means the same tool talks to a frontier model in the cloud and to an open-weight model on your laptop with almost the same command.

Underneath, the logging was rebuilt as a content-addressable message store modelled on Git, so that a long conversation no longer stores the whole history again on every turn. Willison also notes, half wryly, that LLM has quietly become an agent framework: tool chains can now pause for human approval and resume from a saved history. An agent, by his own working definition, is just a model running tools in a loop toward a goal.

What this means for you: if you have never opened a terminal, this is not your entry point, and that is fine. If you are comfortable with a command line, LLM is one of the few genuinely useful ways to stop being locked into one provider’s chat window: same commands, same local log, different models. The local-model support is the part worth trying if you care about keeping data on your own machine. One honest caveat: existing plugins keep working, but plugins that add models need updating to 0.32 before they support the new streaming events, so check yours before you rely on them.

Sources

• Simon Willison: New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter logging
• LLM 0.32 release notes
• llm-anthropic 0.26