Developer jaehongpark-agent published a GitHub Issues post on March 16, 2026, pushing back on two common dismissals of Model Context Protocol: "CLI is enough" and "skills make MCP unnecessary." The post lays out a three-way comparison that treats all three as complementary, each with a distinct cost profile.
CLI tools put the documentation burden on the client side. Well-known tools like the GitHub CLI (gh) are embedded in LLM training data, so their integration cost approaches zero. Lesser-known tools require the agent or developer to supply API documentation and configuration — that burden compounds quickly when an agent needs to connect to dozens of services.
Skills require downloading or writing playbooks. They work well for repeatable workflows but don't scale cleanly across a broad service catalog.
MCP shifts the integration burden to the server side entirely. Once an MCP server exists for a service, any compatible agent connects with a single setup step. The trade-off is straightforward: someone has to build and maintain that server. For widely-used services, that cost gets amortized across every agent that connects — which is the crux of the "build once, connect many" argument.
The post's practical recommendation maps cleanly: CLI for well-known tools, skills for repeatable workflows, MCP for scalable service connectivity.
The GitHub MCP Registry — GitHub's own directory of MCP servers, launched as part of its developer platform — gives the framework some institutional weight. With a central registry, developers can discover and connect to MCP servers without tracking them down individually, which directly addresses the server-maintenance objection the post raises. How fast that registry fills out with quality servers—and <a href="/news/2026-03-14-optimizing-web-content-for-ai-agents-via-http-content-negotiation">how well services optimize their interfaces</a>—will determine whether MCP's theoretical cost advantage holds in practice.