LLM 0.32 adds reasoning traces, server tools, new logging

Simon Willison released LLM 0.32 on August 4, 2026, marking the most significant update to the project since its launch. The release centers on five areas: visible reasoning traces, server-side tools, a redesigned logging system, new models, and an overhauled Python API.

For command-line users, running LLM against reasoning models now displays their reasoning traces to standard error (the intermediate "thinking" steps), so they don't clutter the standard output piped to other tools. Users can suppress this with the -R/--hide-reasoning flag. The CLI now includes GPT-5.6 out of the box, with GPT-5.6 Luna set as the default for the llm "prompt" command. More significantly, LLM can invoke server-side tools provided by cloud platforms. OpenAI exposes CodeInterpreter and WebSearch as server-side tools; users can now run prompts like `llm --tool CodeInterpreter 'Show current python and SQLite versions'` without managing the execution environment themselves. The llm-anthropic plugin (version 0.26) adds WebSearch, WebFetch, CodeExecution, and AnthropicMCP, the last of which allows Anthropic's Claude to execute Model Context Protocol calls against external services (for example, querying a Datasette instance) within a single API interaction.

The Python API received a major redesign. Previously, developers had to create a conversation object and send messages one at a time—an abstraction that obscured how language models actually work, where each request includes the full message history. The new release introduces model.prompt(messages=[]) to work with complete message sequences directly, accepting helper functions like system(), user(), and assistant() to construct the message list. Additionally, LLM now exposes stream_events() to handle the diverse outputs modern models produce: reasoning text, output strings, tool calls, and image attachments. Developers can iterate over events and handle each type separately—for instance, printing reasoning in brackets while streaming regular text to stdout.

These changes enabled a new plugin, llm-chat-completions-server, which implements the OpenAI chat completions API as a local server. Users can install it, run `llm chat-completions-server --port 9000`, and then invoke prompts against that server using the new llm openai endpoint command. This approach pairs well with the llm openai endpoint tool, which lets users run one-off prompts against any OpenAI-compatible endpoint (such as Gemma 4 12B running locally in LM Studio) without persisting logs.

Logging was redesigned around a content-addressable message store inspired by Git. Instead of repeating full message JSON on every turn, the system stores unique messages by hash and references them by pointer, sharply reducing storage overhead for long conversation histories. The llm logs and llm logs --json commands automatically convert this internal format back into standard JSON for easy consumption.

Willison updated several plugins: llm-anthropic now supports the Claude 5 family plus the new server-side tools, and llm-gemini, llm-openrouter, and llm-mistral updates are forthcoming. Existing plugins remain compatible, though model-providing plugins must upgrade to 0.32 to use the streaming events system. Documentation includes guides for implementing plugins with Structured messages and streaming events.

Willison noted that the release blurs the line between a CLI toolkit and an agent framework. Many of the lower-level tool changes were driven by the needs of Datasette Agent, a system that runs tools in a loop to achieve goals. The new ability to pause tool chains for human approval and resume from stored message history supports this use case. With a CLI that can mix models, tools from different sources, and a Python library powerful enough to build systems like Datasette Agent and llm-coding-agent, LLM is "beginning to look very agent-shaped." Willison suggested the next version might bake the agent concept into the core library, though he is still exploring what that would entail.