LLM 0.32: Revolutionizing Reasoning Traces and Advanced Tools

Launch of LLM 0.32: A Major Update

Version 0.32 of LLM has been unveiled, marking a significant milestone since the project's inception. This update introduces innovative features such as support for visible reasoning traces, server-side tools, and a reorganization of SQLite logs. Additionally, it integrates new models and enhanced functionalities through the OpenAI Responses API. A new version of the llm-anthropic plugin has also been released, bringing notable improvements.

Enhancements for LLM CLI Users

LLM CLI users can now display the reasoning traces of models directly in the standard error output. This allows for understanding the models' thought processes without these details appearing in the standard output. For those who prefer not to see these traces, an option -R/--hide-reasoning is available to disable this feature.

Version 0.32 supports the GPT-5.6 model family, with the GPT-5.6 Luna model as the new default for the llm "prompt" command. This model stands out for its efficiency and enhanced performance.

LLM now allows the use of server-side tools from various providers. OpenAI offers a code execution environment as a server-side tool, enabling LLM to execute prompts that leverage it, as illustrated by the following example:

llm --tool CodeInterpreter 'Show current python and SQLite versions'

OpenAI also provides a Web Search tool, expanding the capabilities for interacting with online data.

The llm-anthropic plugin enriches the user experience with tools such as WebSearch, WebFetch, CodeExecution, and AnthropicMCP. For example, the use of the latter tool is illustrated by the following command:

llm -m [claude](/outil/claude)-sonnet-5 -T 'AnthropicMCP("https://datasette.simonwillison.net/-/mcp")' \
'how many rows in the blog_blogmark table?'

A new endpoint command, llm openai, allows executing prompts on any OpenAI-compatible endpoint in a single line, without logging the calls, making it a convenient tool for one-off queries.

Innovations in the Python API

The Python API of LLM has been simplified. Previously, it required creating a conversation and sending messages one by one. Now, a model.prompt(messages=[]) parameter allows sending a set of messages in a single request, better reflecting the nature of LLMs where each request contains the complete history of previous messages.

from llm import user, assistant, system
model = llm.get_model("gpt-5.6-luna")
response = model.prompt(messages=[
    system("You are a helpful pirate."),
    user("What is the capital of France?"),
    assistant("Paris, matey."),
    user("And Germany?")
])
print(response.text())

Version 0.32 also allows handling more complex responses, including reasoning text, output strings, tool calls, and even image attachments. For example:

for event in model.prompt("Explain cats").stream_events():
    if event.type == "reasoning":
        print(f"[thinking] {event.chunk}", end="", flush=True)
    elif event.type == "text":
        print(event.chunk, end="", flush=True)
    print(f"Other event: {event}")

These features enable the implementation of a robust semi-standard OpenAI chat completions API, now available as the llm-chat-completions-server plugin:

llm install llm-chat-completions-server
llm chat-completions-server --port 9000
# The server is now running at http://127.0.0.1:9000/v1

Users can execute prompts via this server using the new llm openai endpoint command:

llm openai endpoint http://127.0.0.1:9000/v1 'hello' -m gpt-5.4-mini

Advanced Logging Management

One of the major challenges of this API is logging. To avoid duplicating JSON data with each request, a new content-addressable message storage system has been introduced, inspired by Git. The llm logs and llm logs --json commands have been updated to convert this format into an easily consumable format.

Evolution of Plugins and Agent Framework

Version 0.32 includes detailed release notes, also covering preliminary versions such as 0.32rc2, 0.32rc, 0.32a3, 0.32a2, and 0.32a0. All existing LLM plugins should continue to function, but those providing additional models will need to be updated to take advantage of the new streaming event system. A guide is available to assist in implementing plugins with structured messages and streaming events.

Some plugins have already been updated, including:

• llm-anthropic 0.26, which supports the Claude 5 model family and includes server-side tools like WebSearch, WebFetch, and CodeExecution.
• llm-gemini, llm-openrouter, and llm-mistral are nearing completion, with upcoming releases.

LLM as an Agent Framework

The changes made to low-level tools in this version have been driven by the needs of Datasette Agent. The term "agent" has gained clarity, and LLM is moving towards a definition where an agent executes tools in a loop to achieve a goal. Tool chains can now be interrupted for human approval and resume from a stored message history, two essential features for Datasette Agent.

LLM is beginning to take the form of a true agent, capable of combining various tools and models into a single command line. It includes a Python library powerful enough to build systems like Datasette Agent and llm-coding-agent. The next version may further integrate the concept of an agent into the main library, although the exact form remains to be defined.