Designing MCP for the Age of AI Agents

The redesigned Harness MCP server v2 offers an efficient and scalable MCP-compatible interface for AI agents, reducing tool count from over 130 to just 11 while minimizing context consumption from 26% to 1.6% in a 200K-token window. By adopting a registry-based dispatch model, it supports 125+ resource types without expanding the tool vocabulary, allowing the LLM to focus on reasoning rather than serving as a routing layer. This architecture optimizes agent performance by eliminating excessive context overhead, enabling more effective execution of complex workflows within developer environments. The v2 server also includes built-in safety controls such as confirmation for writes, fail-closed deletes, and a read-only mode, ensuring secure and reliable operations. Through the integration of Harness Skills, the server supports guided workflows that enhance usability and efficiency, making it a robust solution for deploying AI-driven DevOps processes across platforms like Cursor and Claude Code while maintaining compatibility with various MCP clients.