LM Studio AI Provider Plugin

The LM Studio plugin connects Ever Works to a local or remote LM Studio server for self-hosted AI inference. It extends BaseAiProvider and uses the shared AiOperations layer that wraps LangChain under the hood.

Source: packages/plugins/lm-studio/src/lm-studio.plugin.ts

Overview

Property	Value
Plugin ID	lm-studio
Category	ai-provider
Capabilities	ai-provider
Version	1.0.0
Configuration Mode	user-required
Provider Type	lm-studio
Auto-enable	No
Built-in	Yes
Visibility	public

LM Studio runs open-source models (Llama, Qwen, Mistral, Gemma, …) on your own machine and exposes them through an OpenAI-compatible API. Because the server is local, your data never leaves your infrastructure, and there are no per-token costs.

Architecture

The plugin talks to LM Studio through its OpenAI-compatible /v1 API endpoint, so any model loaded in LM Studio works without additional configuration.

Configuration

Settings Schema

Setting	Type	Default	Scope	Description
baseUrl	string	—	user	Address of the LM Studio server (e.g. http://localhost:1234/v1)
apiKey	string	lm-studio	user	Only needed for an auth proxy in front of LM Studio (stored encrypted)
defaultModel	string	—	global	Used for all AI tasks unless a tier-specific model is set
simpleModel	string	—	global	Handles tags, short descriptions, and quick classifications
mediumModel	string	—	global	Handles listings, summaries, and content reformatting
complexModel	string	—	global	Handles full page generation and multi-step analysis
embeddingModel	string	—	global	Model for semantic-search embeddings (only needed for KB search)
temperature	number	0.7	hidden	Controls output randomness (0 = deterministic, 2 = creative)
maxTokens	number	4096	hidden	Maximum length of each AI-generated response

Model fields use the x-widget: model-select extension, which renders a model dropdown in the dashboard UI populated by calling listModels() against the configured server. Unlike cloud providers, the model fields ship without a hardcoded default — LM Studio serves whatever model you have loaded, so you select it after the connection succeeds.

Required Fields

• baseUrl — the LM Studio server address
• defaultModel — at least one model must be selected

Model Capabilities

getCapabilities(): AiModelCapabilities {
    return {
        supportsStructuredOutput: true,
        supportsStreaming: true,
        supportsToolCalling: true,
        supportsVision: true,
        maxContextLength: 128000
    };
}

These are advisory hints — actual support depends on the model you load in LM Studio.

Lifecycle

Loading

async onLoad(context: PluginContext): Promise<void> {
    await super.onLoad(context);
    this.aiOps = new AiOperations({
        apiKey: 'lm-studio',
        model: this.getDefaultModelId(),
        baseURL: 'http://localhost:1234/v1',
        temperature: 0.7,
        maxTokens: 4096,
        providerType: 'lm-studio'
    });
}

On load the plugin creates an AiOperations instance with default values. When a request arrives, resolveConfig() merges the user's saved settings on top of these defaults before executing.

Availability Check

isAvailable() calls AiOperations.testConnection() with the resolved configuration to verify the LM Studio server is reachable. The setup UI uses this (via validateConnection()) to detect connectivity before saving.

Getting Started

1. Install LM Studio from lmstudio.ai and download at least one model.
2. Start the Local Server in LM Studio (Developer tab) — it listens on http://localhost:1234 by default.
3. Enable the LM Studio plugin in Settings → Plugins.
4. Set the LM Studio Server URL to your instance address (include the /v1 suffix).
5. Select your preferred model for each task complexity tier.

Troubleshooting

Issue	Cause	Solution
Plugin shows unavailable	Local server not started	Open LM Studio → Developer → Start Server
No models in dropdown	No model loaded	Load a model in the LM Studio app
Connection refused	Wrong base URL	Verify the URL includes /v1 (e.g., http://localhost:1234/v1)
Slow responses	Model too large for hardware	Use a smaller / quantized model or increase available RAM/VRAM