DeepSeek Harness setup and Claude Code comparison

This walkthrough installs DeepSeek Harness, an open source coding assistant, and configures it through a local web dashboard. The presenter skips the initial DeepSeek API key prompt and adds a GLM provider in settings. The tutorial separates the model from the harness that manages tools, context, permissions and execution.

For viewers who want to run models locally, the presenter explains how to add an Ollama endpoint as a custom provider. That path is not demonstrated because the presenter's hardware cannot run the intended setup. The working example uses a paid GLM subscription. The speaker distinguishes the free harness from model costs and warns that free hosted models have usage caps.

The hands-on work creates a plugin for completion notifications and token costs, then adds an on-screen notification bar. The generated interface needs fixes for unreadable text and a broken counter. The trajectory view exposes prompts, injected context, tool payloads, results and timing. A separate interruption test shows a task resuming with its recorded context.

The presenter describes the software as a developer preview with a release candidate and expected breaking changes. Claude Code and Codex subagent packages receive particular attention: calls start fresh processes, cannot resume conversations, require preset activation and need credentials passed explicitly. Connecting Anthropic as a model provider uses an API key rather than a Claude subscription.