When developing agents using DeepSeek Harness (DSH), running large models locally usually requires manually configuring an adapter or adjusting code. This not only increases the complexity of environment setup but may also involve unnecessary cloud API calls. The dsh-plugin-lmstudio plugin is designed to simplify this process, mapping LM Studio’s local inference capabilities directly to DSH’s chat model layer.

What Is This

This is a DSH plugin maintained by starlightzfy. It injects an OpenAI-compatible route so that DSH’s model selector can identify and use local models loaded in LM Studio, completing conversations without consuming cloud API quota.

Core Features

  • Use an LM Studio model as the DSH chat model.
  • No cloud API quota required.
  • Communication is sent only to 127.0.0.1:1234.
  • Provides a default lm-studio provider route.
  • No API key required (for the local server).

Installation and Enabling

  1. Ensure DSH CLI and pnpm are installed.
  2. Run the installation command:
    dsh plugin --profile web add dsh-plugin-lmstudio
  1. After installation, restart the DSH Web application (run dsh web) and refresh the browser page (usually http://127.0.0.1:3080).

Configuration and Usage

  1. Start the service and load a model: In LM Studio, go to Developer mode (Developer -> Start Server), default port is 1234, and make sure at least one model is loaded.
  2. Get the model ID: Run curl http://127.0.0.1:1234/v1/models in the terminal, and check the id field in the returned data (usually formatted as owner/name).
  3. Configure the model: In the DSH interface, click the “ask the endpoint” button under “Settings → Models” to automatically discover models, or edit the configuration file directly.
  4. Edit the configuration file: The plugin’s built-in lm-studio route is the base layer, and can be overridden in $DSH_HOME/settings.yaml.

Example configuration:

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

The configuration file supports hot reloading, so there is no need to restart dsh web.

Notes

  • You must start the server in LM Studio (default port 1234) and load at least one model.
  • The plugin runs with the permissions of the current dsh process; make sure the environment is secure.
  • If LM Studio has key authentication enabled, add apiKeyEnv in the DSH settings layer or add Authorization to the request headers.
  • The plugin license is MIT.

Conclusion

After the above configuration, DSH’s chat model layer can connect to local models. This plugin is suitable for developers who need to run models in isolated local environments and avoid cloud API quota limits. The plugin repository and directory links are as follows:
- GitHub repository
- Plugin directory