MCP Is Not Enough (mukulsingh105.github.io)

Anthropic's Model Context Protocol (MCP) is gaining traction as a standardized method for AI model integrations, akin to the universal USB-C connector. While MCP excels in facilitating simple, single-step tool calls across models like Claude and ChatGPT, it struggles with more complex, multi-step workflows crucial for real-world applications such as order fulfillment and CRM coordination. This limitation stems from requiring multiple round trips for each step, resulting in significant drop-offs in accuracy and increased potential for errors, as evidenced by findings from the Berkeley Function Calling Leaderboard. The author argues for the development of domain-specific languages (DSLs) to enhance multi-step task execution. DSLs allow models to generate coherent code in a single pass, retaining context and reducing the risk of errors related to parameter handling and nested JSON structures. By replacing the JSON format with typed, structured language, the process becomes more streamlined and less prone to hallucinations. Experiments suggest that using a DSL can drastically cut down on token use and processing time, making it a superior alternative to traditional JSON tool calling for complex workflows. The proposal emphasizes that combining MCP's connectivity with a DSL can optimize AI production processes significantly.