In early 2026, developers sought a single gateway to let LLM agents interact with systems. Traditional MCP servers hand agents a menu of narrow tools—create_issue, list_pull_requests, and so on. The approach feels tidy, but the article argues that a single, flexible CLI outperforms hundreds of specialized wrappers for most real workloads and reduce context bloat.
MCP design bundles each service into a pile of typed tools, forcing agents to pick the right one. In contrast, a CLI exposes a single expressive command; the model learns to parse help text, assemble pipelines, and retry flag errors. Recent models can read --help and compose bash scripts without external scaffolding for efficient automation.
Extending to several databases magnifies the cost. An MCP agent must load a separate tool schema per environment, inflating the context window by dozens of tokens. A single CLI binary, however, can switch profiles in a loop, keeping the token budget flat regardless of how many instances the agent interrogates across production and staging environments.
Pipeline composition turns the CLI into a low‑level orchestration tool. A two‑step query that pulls counterparties and then aggregates edges can be executed as a single pipe, keeping intermediate data out of the model’s context. The result is a leaner, faster interaction that still lets agents wield complex workflows across various domains while maintaining security.
Source: Towards Data Science · Summarized by HeadlinesBriefing