DeepSeek Harness: Qwen 3.8 setup with Ollama and plugins

This tutorial sets up DeepSeek Harness as a customizable coding assistant and connects it to Qwen 3.8 through Ollama. The presenter describes an open source system whose interface, agent loop and tools are plugins that users can disable or extend.

The walkthrough covers installation, browser access through dsh web, workspace permissions and agent modes. A DeepSeek API key is optional. The presenter demonstrates DeepSeek V4 Pro and explains that other providers and local models can also connect. Minimal mode limits the agent to a terminal and file editor.

For the local LLM setup, the presenter recommends about 16 GB of VRAM for the Qwen 3.8 model used. Ollama connects as a custom provider at localhost port 11434 with /v1 and the OpenAI completions protocol. A custom preset removes sub-agents and the Ralph loop to reduce the tools supplied in context. This reflects the presenter's advice for the demonstrated setup, rather than a universal limit on local models.

The plugin section installs a browser sidebar, then uses creator mode to build a confetti effect. For that creation step, the presenter switches from the local model to the Pro model and enables full access. The tutorial also explains that a dynamic plugin runs in memory: users must generate permanent plugin files before closing the harness if they want to keep or distribute it.