Why we deleted (most of) our MCP tools

The evolution of AI tools has shifted from using Model-Contextualized Protocols (MCPs) to a more efficient system of skills and subagents, improving the functionality and token efficiency of AI assistants. While MCPs were initially crucial for integrating AI with developer environments by offering tools like project graphs and task pipelines, advancements such as the Claude Code have enabled AI agents to execute commands autonomously, reducing the need for MCPs. Skills, which provide domain-specific knowledge and instructions, have proven to be more effective, allowing agents to utilize command-line interfaces like Nx CLI directly and make informed decisions about tasks such as generator usage and workspace exploration. Despite the decline in MCP usage, it remains valuable for authenticated API interactions and processes that require connectivity beyond the agent's autonomous capabilities. The ongoing development focuses on refining skills, introducing new subagents, and leveraging MCP for tasks that are not easily replicable through skills alone, as demonstrated in benchmarks comparing agent performance with and without these enhancements.