DeepSeek Harness: Customizability vs. Security Risks

Architecture and Customization

DeepSeek Harness distinguishes itself from existing developer harnesses by treating the entire application as a modular plugin system built on the Cordis engine. Unlike Claude Code or Codeex, which provide fixed interfaces with limited hooks, DeepSeek Harness allows users to modify the core agent loop and UI components directly. Users can enter "Creator Mode" to prompt the harness to generate custom plugins that alter its own behavior, such as modifying the UI color scheme or adding specific workflow sliders. The system provides granular visibility into the execution process via a "Trajectory" view, which exposes system prompts, token usage, cache hit percentages, and raw tool-call sources for debugging.

Operational Risks and Security

While the extensibility is a significant differentiator, the current developer preview version presents substantial security concerns. Every installed plugin currently operates with full shell access and unrestricted read/write permissions to the user's file system. Because the ecosystem relies on community-contributed plugins, there is a high risk of malicious code execution or credential theft if users install untrusted plugins. The author recommends auditing plugin source code—potentially using another LLM like Claude Code—before execution. Additionally, while the core harness is open-source and supports local model execution via providers like OpenRouter, specific features like web search require manual configuration or the installation of third-party plugins if the user does not utilize the native DeepSeek API.

Workflow Modes

• Standard Mode: Executes tool calls sequentially, mimicking standard agentic workflows.
• PTC Mode: Optimizes complex tasks by batching tool calls into a single script to reduce context bloat.
• Minimal Mode: Optimized for simple, low-overhead tasks.
• Creator Mode: Enables the generation and integration of custom plugins for the harness environment.