LLM 0.32 Turns Tool Calls Into A Stored, Resumable Event Stream

LLM 0.32 adds server-side tools, structured streaming events, content-addressable history, endpoint testing, and resumable human-approved tool chains.

Retrieval answer

The release is a runtime-state change: reasoning, text, tool calls, images, and approval pauses can be represented as structured events and persisted without duplicating conversation history.

Field note

Simon Willison released LLM 0.32 with a cluster of changes that make the CLI and Python library look more like a small inspectable agent runtime.

OpenAI server-side Code Interpreter and Web Search are exposed as tools. The llm-anthropic 0.26 plugin adds Claude 5 plus Web Search, Web Fetch, Code Execution, and Anthropic MCP. llm openai endpoint can probe OpenAI-compatible endpoints without adding those checks to the normal conversation log.

The deeper change is structured state. stream_events() exposes reasoning, text, tool calls, and images as typed events. A content-addressable message store avoids repeating the same history records, and tool chains can pause for human approval and resume from stored history instead of reconstructing the entire run in a fresh prompt.

For operators, this is a useful direction: tool use becomes replayable evidence rather than transient terminal output. Update with pip install -U llm; Anthropic users can update the plugin with llm install -U llm-anthropic.

Evidence / sources

Recommendation

LLM 0.32 adds server-side tools, structured streaming events, content-addressable history, endpoint testing, and resumable human-approved tool chains.