llm-chat-completions-server 0.1a0 リリースllm-chat-completions-server 0.1a0

Simon Willison has published llm-chat-completions-server 0.1a0, a new plugin for his LLM command-line tool and Python library that spins up a local API server compatible with OpenAI's Chat Completions format. The release matters because it lets any software written to talk to OpenAI's API point instead at a locally running instance of LLM, opening the door to using local models, alternative providers, or logged conversations through a widely supported interface.

Once installed and started, the plugin listens on a local address — the example in the announcement uses http://localhost:8002/v1/chat/completions — and accepts the same JSON request shape that OpenAI's endpoint expects. A basic call can be made with a simple curl command, sending a list of messages and receiving a model response in the familiar Chat Completions structure. Because the path and payload mirror OpenAI's, existing client libraries and tools should be able to use the server with little more than a changed base URL.

The feature that distinguishes this plugin is its handling of multi-turn conversations. Willison notes that a key goal of the content-addressable logs introduced in LLM 0.32rc1 was to support requests in which each incoming message extends the previous conversation. In the standard Chat Completions model, every request is stateless and carries the full message history. By addressing stored messages by the hash of their content, the server appears to be able to recognize when a new request shares a prefix with an earlier exchange, allowing it to treat the interaction as a continuation of an existing thread rather than an unrelated new log. This is what enables the server to record stateful, multi-turn chats while still speaking the stateless OpenAI protocol.

For readers unfamiliar with it, LLM is an open-source tool that provides a consistent way to run prompts against many different models from the command line or from Python. It ships with support for hosted APIs and gains access to local models through a plugin ecosystem that includes options such as llm-ollama, llm-mlx, llm-gpt4all and llm-llama-cpp. Conversations and responses are stored in a SQLite database, and the logging system has long been a core part of the tool. The move to content-addressable logs in 0.32rc1 is the groundwork that this server plugin builds on.

The broader context is the rise of the OpenAI Chat Completions API as a de facto standard. A large number of applications, agent frameworks and developer tools default to that request format, and many alternative backends — from local inference servers like llama