Introduction¶
DeepSeek Harness (DSH) adopts a plugin-based architecture. When running large models locally, it is usually necessary to separately manage the llama-server process, manually configure the port and model path, and switch between different interfaces. The dsh-llamacpp-bridge plugin is designed to eliminate these intermediate steps, integrating local llama.cpp directly into DSH’s model provisioning system so that it can be managed and called like a cloud model.
Plugin Overview¶
dsh-llamacpp-bridge is a plugin that turns local llama.cpp (with llama-server) into a first-class model provider for DSH.
- Maintainer: b8yg7vjstj-ctrl
- License: MIT
- Core Value: Unifies local model management and provides process control, catalog synchronization, automatic pairing, and visual monitoring.
Core Features¶
This plugin provides the following capabilities:
1. Provider Role: Registers local llama.cpp as a DSH-compatible model provider.
2. Mode Switching: Supports switching between single-model mode and Router mode.
3. Catalog Synchronization: Reads the model list directly from a specified folder, with automatic rescanning and change detection.
4. Model Switching: Stops the old model before loading a new model to avoid residual VRAM.
5. Vision Pairing: Supports automatic pairing of mmproj visual projection files.
6. Sidebar Monitoring: Provides a terminal-output monitoring panel showing status, model-switching buttons, and scrolling logs.
7. Graceful Exit: Supports graceful process termination via Ctrl+C.
8. Hot Application: Settings changes take effect immediately without a restart.
Installation and Activation¶
Before installation, ensure the environment meets the requirements: DSH (web profile), Node.js 20+, and a local llama.cpp installation with llama-server.
dsh plugin --profile web add ./dsh-llamacpp-bridge-1.5.0.tgz
After installation, you must restart the DSH host and hard refresh the browser (Cmd/Ctrl + Shift + R) to ensure the plugin client scripts load correctly.
Typical Usage¶
- Startup and Configuration: After restarting DSH, locate “llama.cpp Terminal Output Monitoring” in the “New Session” area of the sidebar. Go to Settings → llama.cpp to manually specify or confirm the path to
llama-serverand the model directory. - Using a Model: Select a local model in the chat interface and send a message. The plugin automatically starts
llama-serverin the background and forwards requests once it is ready. - Monitoring and Switching: Click the sidebar panel to view the current status, model-switching buttons, and scrolling logs. Use the panel buttons to switch models or refresh the list directly.
Settings Description¶
In Settings → llama.cpp, you can configure the following fields:
| Field | Default | Description |
|---|---|---|
displayName |
Local llama.cpp (bridge) |
Name displayed in the model selector |
executable |
Empty | Path to the llama-server executable |
modelsDir |
Empty | Directory containing model files |
host / port |
127.0.0.1 / 8080 |
Service listening host and port |
strategy |
auto |
Runtime strategy: auto / single / router |
contextLength |
— | Context length passed to -c |
autoContext |
false |
Whether to infer the context size based on available VRAM |
gpuLayers |
-1 |
Number of GPU layers passed to -ngl |
mmprojSuffix |
-mmproj.gguf |
Suffix naming convention for visual projection files |
mmprojOverrides |
Empty | Manually bound mapping between models and projection files |
additionalArgs |
Empty | Additional arguments passed to llama-server |
startTimeoutMs |
120000 |
Startup readiness timeout (milliseconds) |
debug |
false |
Whether to enable debug logs |
Notes¶
- Dependency Environment: Depends on DSH web profile and Node.js 20+.
- Restart Mechanism: After installation, you must restart the DSH host and hard refresh the browser, otherwise errors may occur.
- Uninstall Command:
dsh plugin --profile web remove dsh-llamacpp-bridge.
Ecosystem Context¶
The DSH ecosystem emphasizes that everything is a plugin. Community catalogs (such as SkillHub) provide plugin indexes and have no affiliation with the official project.
Conclusion¶
Through a unified interface and monitoring panel, this plugin simplifies the onboarding process for local LLMs. It is suitable for developers who need to call local models reliably within DSH.
Project Links:
* GitHub
* Plugin Directory
