MD Feedback: Guiding AI Agents with Markdown-Based Review in VS Code

MD Feedback introduces a structured methodology for human-in-the-loop oversight in AI agent development, specifically targeting the review of AI plans written in Markdown. Positioned as a VS Code extension and an MCP (Markdown Control Protocol) server, the tool aims to bridge the gap between human intent and AI agent execution. Its core premise is that explicit markdown plans should guide AI agents, with annotations serving as direct instructions or queries for refinement. This approach ensures that human developers retain critical control and context throughout the AI coding loop, from initial plan creation to final implementation.

The workflow is designed for efficiency: developers write a plan in Markdown, use the MD Feedback sidebar in VS Code to highlight, fix, or question specific sections (via simple keyboard shortcuts), and the AI agent then follows the plan, applying these annotation memos. This iterative process allows for continuous review and adjustment, with quality gates automatically tracking completion and session handoffs preserving context across multiple AI agent sessions. The tool supports three distinct annotation types—Highlight (for reading marks), Fix (for required changes), and Question (for clarifications)—providing a nuanced communication channel between human and AI.

Technical features underscore its robustness and interoperability. MD Feedback integrates with 28 MCP tools for direct agent interaction and offers export/share capabilities to 11 popular AI tools, including Claude Code, Cursor, and Gemini, enhancing its versatility within diverse development environments. Annotations are stored as portable HTML comments within the Markdown files, ensuring compatibility with any Markdown renderer and seamless integration with version control systems like Git. Furthermore, the tool incorporates critical safety features, such as blocking writes to sensitive files (`.env`, `credentials`), implementing concurrent safety for multiple AI operations, and providing detailed inline diffs for proposed AI changes, empowering developers to approve or reject modifications with confidence.

By formalizing the review process for AI-generated plans, MD Feedback addresses a growing need for reliable human oversight in autonomous AI development. It transforms what could be a chaotic, opaque process into a transparent, controlled, and collaborative one. This not only helps in preventing errors and ensuring code quality but also fosters a more effective partnership between human developers and AI agents, ultimately accelerating the development lifecycle while maintaining human accountability and strategic direction.