Why I'm building my own CLIs for agents (martinalderson.com)

In a recent discussion, a developer shared their pivot from using Model-Contextualized Pipelines (MCP) to building custom Command Line Interfaces (CLIs) for managing Large Language Model (LLM) interactions. The developer noted that while MCP aims to seamlessly connect various data sources to LLMs, it suffers from significant token consumption problems, particularly due to context length constraints. With the complexity of attention mechanisms, handling sessions with extensive token counts becomes resource-intensive and inefficient. This inefficiency was exemplified when using the Linear MCP, which consumed vast amounts of tokens just for tool definitions, limiting the effective context available for meaningful interactions. By creating their own CLIs, the developer found a solution that significantly reduced token usage while maintaining functionality. For instance, they demonstrated how simple CLI commands for tasks like creating or updating issues in project management tools used only a fraction of the tokens compared to MCPs. This approach not only streamlines operations but also enhances collaborative coding efforts by ensuring all team members are aligned on CLI usage. The developer sees this trend as an evolution in LLM applications, emphasizing the importance of CLIs in reducing overhead and improving user experience, especially as the reliance on complex MCPs continues to pose challenges. As LLM technology progresses, this method could signal a shift towards more efficient, tailored solutions in the AI/ML landscape.