@namzu/lmstudio

LM Studio LLM provider for @namzu/sdk. Thin wrapper around the official @lmstudio/sdk that conforms to the LLMProvider contract (chat + streaming). Run any open-weights model locally through LM Studio's GUI-managed server.

Install

pnpm add @namzu/sdk @namzu/lmstudio

@namzu/lmstudio declares @namzu/sdk as a peer dependency. Install both.

Prerequisites

Requires LM Studio running as a local server (default http://localhost:1234, transported over WebSocket by the official SDK). Load a model via LM Studio's UI or the CLI:

lms load <model>

Start the local server from LM Studio's Developer tab (or lms server start).

Usage

import { ProviderRegistry } from '@namzu/sdk'
import { registerLMStudio } from '@namzu/lmstudio'

// Register once at app startup.
registerLMStudio()

// Fully typed via module augmentation: ProviderConfigRegistry['lmstudio'].
const { provider, capabilities } = ProviderRegistry.create({
  type: 'lmstudio',
  host: 'http://localhost:1234', // optional — defaults to LM Studio's auto-detection or LMSTUDIO_HOST env var. Accepts http(s) (auto-converted to ws(s) for the SDK).
  model: 'llama-3.2-1b-instruct', // optional default — can also be passed per-call
})

const response = await provider.chat({
  model: 'llama-3.2-1b-instruct',
  messages: [{ role: 'user', content: 'Hello' }],
})

Streaming:

for await (const chunk of provider.chatStream({
  model: 'llama-3.2-1b-instruct',
  messages: [{ role: 'user', content: 'Tell me a story' }],
})) {
  if (chunk.delta.content) process.stdout.write(chunk.delta.content)
}

Configuration

| Field | Description | Default | |-----------|-----------------------------------------------------------------------------|------------------------------| | host | LM Studio server URL (http://… accepted, converted to ws://…) | Auto-detected by @lmstudio/sdk or LMSTUDIO_HOST | | model | Default model identifier (overridable per-call via ChatCompletionParams.model) | — | | timeout | Request timeout in ms | SDK default |

The model identifier must match a model already loaded in LM Studio. Use lms ls or the LM Studio UI to discover loaded models; provider.listModels() returns loaded models via the SDK's listLoaded().

Transport

Uses the official @lmstudio/sdk (WebSocket). This unlocks richer local-model features than the OpenAI-compat HTTP endpoint — the SDK is built for local-LLM workflows (loading/unloading, configuration, speculative decoding). If you prefer zero-dep HTTP instead, use @namzu/http with baseURL: 'http://<host>/v1' and dialect: 'openai'.

Capabilities

import { LMSTUDIO_CAPABILITIES } from '@namzu/lmstudio'

// {
//   supportsTools: true,
//   supportsStreaming: true,
//   supportsFunctionCalling: true,
// }

Tool-use support depends on the loaded model — not every open-weights model has been tuned for function calling. Refer to the model's card in LM Studio.

Observability

This package ships without observability hooks in 0.1.x. OpenTelemetry span emission and structured logging are roadmapped for the forthcoming @namzu/telemetry package — a separate opt-in dependency that wraps any LLMProvider to emit GenAI semantic-convention spans. Track progress in the Namzu roadmap.

License

FSL-1.1-MIT. Same as @namzu/sdk.