setup

Set up Local Model Workers

The plugin ships the MCP server, but the server needs two things the plugin cannot supply: a provider connection (protected, environment-only) and a default model (a global preference).

Work through the steps in order. Stop as soon as check_health is healthy.

1. Confirm what is already configured

Call check_health and get_config. They need no repository and return redacted values. If health is healthy and a default model is routed, there is nothing to do.

2. Point the server at the user's local model runtime

Connection settings are protected: they live in the process environment only, so repository content and preference files can never weaken them. The plugin reads them from Claude Code's environment.

The reliable place to set them is the env block of the user's Claude Code settings (~/.claude/settings.json), which applies to the terminal, the desktop app, and the IDE extensions alike:

{
  "env": {
    "LMW_PROVIDERS": "[{\"name\":\"lm-studio\",\"type\":\"lm-studio\",\"base_url\":\"http://localhost:1234/v1\",\"allowed_models\":[\"qwen/qwen3.5-9b\"],\"priority\":0}]"
  }
}

LMW_PROVIDERS is required: a JSON array of one to 16 objects with name, type (lm-studio, ollama, vllm, localai), base_url, optional bearer_token, allowed_models, priority (lower wins), and optional tls_verify.

• base_url — absolute http(s) URL ending in /v1, no credentials, query, or fragment. LM Studio serves 1234, Ollama 11434/v1, vLLM and LocalAI 8000/v1.
• allowed_models — model identifiers the server may use. ["*"] means "any served model".
• bearer_token — optional; omitting it selects no authentication, which is the expected posture for a trusted LAN.
• tls_verify — optional. Unset means remote providers verify certificates and loopback does not. Set it to false to accept a remote plain-HTTP endpoint deliberately.

Claude Code must restart to pick up a changed env block.