DeepSeek Harness setup: plugins, tracing and Ollama

DeepSeek Harness is presented as a developer preview with a plugin-based interface and detailed session tracing. The tutorial builds it from source with pnpm install and pnpm run build, then starts the browser interface with pnpm dsh web. Initial setup asks for a DeepSeek API key, but the presenter also configures other providers.

A Flask to-do app task demonstrates the harness as a coding assistant. Its trajectory view exposes the system prompt, loaded context, assistant messages and tool calls, including command payloads, results and timing. The presenter also exports a session log as a ZIP containing JSONL data.

Plugin controls extend to the sidebar itself. In this preview, the presenter disables it through a YAML configuration file and reloads the interface. Creator mode generates a moving cat overlay and a calculator after approval. Both examples have rough edges: the cat animation is basic, and the calculator's button labels are hard to see. The generated plugins are active in the session; the presenter says they can also be persisted.

For local AI use, the tutorial connects an Ollama server on a Dell Pro Max with GB10 and tests a 120-billion-parameter model with a greeting. This confirms a provider connection in the demonstration, rather than establishing its performance on coding tasks.