Introduction

The core design philosophy of DeepSeek Harness (DSH) is “everything is a plugin”. For users who need to deploy large models locally, manually maintaining the llama-server process and its configuration parameters can often be cumbersome. The dsh-local-llm-controller plugin aims to solve this problem by seamlessly integrating a local llama.cpp server into DSH’s settings and model management workflow, enabling one-click start/stop and configuration management.

Plugin Information

  • Name: lbunc/dsh-local-llm-controller
  • Author: Lbunc
  • Category: Model Inference
  • License: MIT
  • Core Value: One-click start/stop a local llama.cpp large model (35B/9B, vision × text × fast/long context) from the “Settings → Plugins” page, with in-card configuration, one-command installation, and automatic registration—ready to use after installation.

Installation and Activation

Using a single command completes installation and automatic registration without manual configuration:

dsh plugin --profile web add dsh-local-llm-controller

After installation, restart DSH Web. The plugin card will automatically appear on the “Settings → Plugins” page.

Core Features

The plugin provides the following capabilities:
1. Card Control: One-click start/stop of llama-server on the settings page, with the status bar displaying running logs.
2. Dual-Slot Management: Supports two model slots, A and B, each with an independently configurable model folder path.
3. Mode Switching: Supports switching between text mode and vision mode.
4. Parameter Presets: Provides eight startup parameter combinations (2 slots × text/vision × fast/long context), with manual editing supported.
5. Automatic Registration: Automatically manages server ports and parameters without manually creating configuration files.
6. Multi-language Adaptation: The card’s language follows DSH Web’s language setting.

Usage Workflow

  1. Configure the Plugin: In the plugin card, fill in the llama.cpp directory (containing llama-server.exe), port (default 55555), and key.
  2. Configure Model Slots: In the “Slot A / Slot B” configuration area, enter the model folder path (including GGUF files and the mmproj file for vision mode), then click Save.
  3. Add to the Model List: After saving the configuration, the GGUF models in the folder will appear as bubbles. Select one model by clicking it, then click “Add to Model List” again.
  4. Edit Parameters: In the “Startup Parameters” section, select or edit one of the eight parameter sets. Base parameters are pre-filled, and the plugin automatically manages parameters such as -m, -a, and --port; no manual addition is required.
  5. Start the Service: At the bottom of the card, select the slot, mode, and parameter preset, then click “Start”.
  6. Chat: After the service starts, select the corresponding local model from the model list at the bottom of a DSH conversation to begin chatting.

Notes

  • Version Compatibility: Compatible only with the DSH RC official branch (currently verified version 0.1.5-rc.1); compatibility with other branches is not guaranteed.
  • Vision Mode: Requires a llama-server build that supports mmproj. Older builds do not support WebP images, so converting them to PNG/JPEG is recommended.
  • Model Switching: After replacing a model file in the same slot, the Provider Key does not update automatically. You must manually delete the old entry on the “Settings → Models” page and then re-add it.
  • Component Dependencies: The plugin does not include the llama-server binary itself and does not handle model downloads; these must be prepared separately.

Ecosystem Background

The DSH philosophy is “everything is a plugin”. The community directory is maintained by SkillHub and is not affiliated with DeepSeek official.

  • Plugin Directory: https://www.skillhub.cn/plugins/Lbunc/dsh-local-llm-controller
  • Source Code Repository: https://github.com/Lbunc/dsh-local-llm-controller