MCP is dead? (www.quandri.io)

The MCP (Model Context Protocol), initially celebrated as a transformative connector for large language models (LLMs) and external tools, has been deemed inefficient and is increasingly viewed as obsolete by developers. The protocol was criticized for consuming excessive context in LLMs—over 10% of the context window being used just for tool definitions—which hampers performance and efficiency. Additionally, benchmarks revealed that MCP calls can be significantly slower compared to direct API access, with some processes being three times slower and utilizing approximately 65 times more tokens than traditional command-line interface (CLI) methods. The significance of this critique lies in its implications for the AI/ML community, particularly for LLM application development. With a shift towards Skills that utilize existing CLIs, developers report substantial reductions in context usage and improved performance, freeing up around 21,000 tokens in daily workflows. This approach promotes greater flexibility, allowing for effective tool use without the cumbersome overhead associated with MCP, while still leveraging the powerful capabilities of LLMs. As the ecosystem evolves, the focus is likely to pivot from broad, all-encompassing solutions like MCP to more tailored methods that prioritize efficiency and user needs, underscoring the importance of practical, user-driven innovation in the field.