dsh-acp-plugin

With the rise of AI agents and tool calls, ensuring each operation adheres to security policies is crucial. The dsh-acp-plugin adds a smart protective layer for DeepSeek Harness, automatically checking policies before tool calls execute to block unauthorized or high-risk actions, thus preventing potential security vulnerabilities. It logs detailed decisions, including the tool called, input preview, reason, and latency, creating a complete audit trail to help users track and verify all activities.

The plugin's core capability lies in its policy engine, which supports three decision modes: allow, deny, and ask. These correspond to direct pass-through, interception, or handoff to user approval, offering flexibility for different scenarios. Whether in interactive sessions or unattended agents, it ensures security without compromise—failing open with loud warnings in interactive mode while failing closed in unattended runs. All decisions are also sent to the Agentic Control Plane console for independent monitoring.

This plugin is ideal for developers and teams using DeepSeek Harness, especially those needing enhanced security control and auditability. Whether you run a single tool or multiple AI agents like Claude Code or Cursor, dsh-acp-plugin integrates seamlessly, applying uniform policy rules across all tool calls to help you easily achieve compliance and risk management in your development workflow.