Choosing a Protocol

LM Studio speaks several APIs on one port, and runs whatever model you have downloaded behind all of them. This gem registers a protocol for each:

API	Endpoints	Protocols
OpenAI	POST /v1/chat/completions, POST /v1/responses, GET /v1/models, POST /v1/embeddings	:chat_completions (default), :responses
Anthropic	POST /v1/messages	:anthropic
Native REST	v0: POST /api/v0/chat/completions, GET /api/v0/models v1: POST /api/v1/chat, GET /api/v1/models, POST /api/v1/models/load, /unload, /download	:native_v0, :native_chat
Decisions	POST /v1/decisions, POST /v1/systemone	:decisions (default for judgments), :system_one

The first three APIs carry the five chat protocols, selected with lms_protocol. The decision protocols serve only judgments and are selected separately with lms_judgment_protocol — see Configuration.

Two choices, not one

The protocol and the model each decide part of the outcome

The protocol decides what is expressible. with_schema is honored on :chat_completions and :native_v0 and silently ignored on :responses and :anthropic — same model, same schema, prose where you expected JSON. Reasoning is spelled none/minimal/low/medium/high/xhigh on the OpenAI endpoints and off/low/medium/high/xhigh/on on the native one. That much is consistent for every model, and this gem papers over what it can.

The model decides whether it honors what the protocol accepted, and no provider code can fix that. LM Studio forwards tool_choice: "required" for anything — qwen3 obeys it, gpt-oss ignores it. It constrains generation with a grammar built from your schema, so every model returns structurally valid JSON — and gpt-oss returns {"name":"analysis","age":0}, which parses cleanly and means nothing.

This is the price of LM Studio's reach: it is all things to all models. Test against the models you actually intend to ship on, rather than assuming a capability carries across either axis.

The chat protocols compared

Stay on the default unless you need something only another column offers.

	:chat_completions	:responses	:anthropic	:native_v0	:native_chat
Streaming	yes	yes	yes	yes	yes
Client-side tools (with_tools)	yes	yes	yes	yes	no — raises; use MCP integrations
Structured output (with_schema)	yes	ignored, silently	ignored, silently	yes	no — raises
Inference stats	no	no	no	yes, via raw	yes, via tokens
LM Studio request extras	yes	yes	yes	yes	sampling settings only — see Native REST API
Server-stored conversations	no	no	no	no	yes
Reasoning control	yes	yes	yes	yes	yes
Reasoning text returned	yes	yes	yes	yes	yes

Checked against LM Studio 0.4.24+1; the :anthropic reasoning rows and the decision endpoints against 0.4.26. The cells that say the server ignores something describe observed behavior, not anything the gem enforces.

Where a protocol ignores something silently rather than raising, its page says so — those are the cases that cost you an afternoon. Example 15_protocol_matrix.rb runs the with_schema row against your own models.

"Reasoning control" means with_thinking reaches the server. Whether a given model reasons at all is a separate question — see Reasoning.

Selecting one

RubyLLM.configure { |config| config.lms_protocol = :anthropic }      # process-wide

RubyLLM.context { |config| config.lms_protocol = :native_v0 }        # one context
       .chat(model: 'qwen/qwen3-4b', provider: :lms)

RubyLLM.chat(model: 'qwen/qwen3-4b', provider: :lms,                 # one chat
             protocol: :native_chat, assume_model_exists: true)

Whatever lms_protocol says, embeddings go to /v1/embeddings and judgments go to a decision endpoint.