I tested DeepSeek Harness for a week. I left with a shipped plugin and $0 in API costs

I tested DeepSeek Harness for a week and left with a shipped plugin and $0 in API costs. Initially, I came to DeepSeek Harness as a reviewer and left as one of its plugin authors. The main feature of the harness is that everything is a plugin, including the settings screen where you pick and configure models. This means that even optional subagents like Codex and Claude Code ship as plugins that can be added or dropped.

All of these plugins are loaded through manifests over a real extension API, which is something I've never experienced with other extensible tools before. During my testing, I encountered a gap in the Harness settings - I wanted to know the cost per million tokens for each model and if a cheaper alternative was available. To address this, I created a plugin called dsh-model-pricing that provides per-1M-token input/output/cache prices for 7,178 priced models across 213 providers.

The plugin also includes a cheapest badge to indicate the lowest-cost option for each model, with any discrepancies flagged as ⚠. The plugin's settings page is integrated into the Harness itself, which is a significant aspect of the testing process. The Harness architecture serves as the product, while the agent performs both the reviewing and building tasks effectively, even when using a free model like Qwen3.8-Flash.

The entire project was built on a free model, and the plugin itself ships with records from a free API promotion, resulting in a total LLM spend of $0 for the project.