What is special about MCP? (jessitron.com)

Model Control Protocol (MCP) is a way to give AI agents access to specific, curated tools (like reading/updating Google Calendar or extracting Figma designs) rather than relying only on generic capabilities (run bash, curl, or a browser). The writeup argues MCPs don’t expand what an agent can ultimately do—an agent with shell access can script API calls—but they add three practical advantages: built-in authentication (OAuth-style, properly attributing actions), a specialized, token-efficient interface (MCP responses can summarize JSON, emit CSV or ASCII art, and avoid verbose payloads), and rapid iteration (MCP providers can change tool formats or add/remove tools without breaking programs). The piece also highlights costs and safety trade-offs: every configured MCP adds instructions to the model context, which can bloat tokens; poorly scoped access risks data leakage or destructive actions that a generic shell could perform; and current MCPs are globally visible to all agent threads. Best practice is to use MCPs when you need curated, authorized access to SaaS/data and to prefer local MCPs or fine-grained per-agent permissions for safety. Future improvements suggested include incremental loading of MCP context and per-subagent visibility so agents only see the tools they actually need.