dsh-plugin-lmstudio

dsh-plugin-lmstudio

Use a model served by LM Studio as a chat model inside the
DeepSeek Harness (DSH) web GUI.

LM Studio exposes an OpenAI-compatible endpoint (http://127.0.0.1:1234/v1 by
default). This plugin seeds DSH's built-in llm-pi-ai adapter with that route,
so the model picker shows your local model — no cloud API quota consumed.

中文说明 · MIT

Requirements

• DSH CLI + pnpm (see the DSH docs)
• LM Studio with Developer → Start Server running (default port 1234)
• At least one model loaded in LM Studio

Install

dsh plugin --profile web add dsh-plugin-lmstudio

The package declares dsh.bundle, so dsh plugin automatically adds it to the
profile's bundle layers. Then restart the web app (dsh web) and refresh
http://127.0.0.1:3080.

Configuration

The plugin ships a default provider route under the lm-studio provider key:

The bundled route is a base layer: anything you write under the
llm-pi-ai: section of $DSH_HOME/settings.yaml merges over it per provider,
so your own models and endpoints win. No API key is required for a local LM
Studio server; if you enable key auth in LM Studio, add apiKeyEnv or an
Authorization header in your settings layer.

Point it at your actual model

1. In LM Studio, load the model you want.
2. Find the exact model id: curl http://127.0.0.1:1234/v1/models and read the
   id field (usually owner/name, e.g. qwen/qwen3.5-9b).
3. Either edit the route in Settings → Models (the "ask the endpoint"
   button discovers models for you), or override it in $DSH_HOME/settings.yaml:

llm-pi-ai:
  providers:
    lm-studio:
      models:
        - id: qwen/qwen3.5-9b
          name: Qwen3.5 9B (LM Studio)
          contextWindow: 8192
          maxTokens: 4096

settings.yaml hot-reloads — no dsh web restart needed.

Usage

Refresh the page, open the model picker above the composer, and choose your
LM Studio model (it appears under "LM Studio (本地)"). Traffic goes to
127.0.0.1:1234 only; your DeepSeek API balance is untouched.

Tip: for plain conversation without any agent tools, create a "chat" agent
preset (an empty agent.cordis.yml) and pick it on the new-session screen —
that applies to every model, local or cloud.

License

MIT