MCP and function calling hand the model the same tool schema

The dispatch loop is where the equivalence gets concrete. You hand the model a tool name, a description and an input_schema. If it decides mid-run that it needs that tool, the response content carries a tool_use object with the name and the arguments it chose, and stop_reason comes back as tool_use ^[6]. Your code branches on that and calls the real function ^[6]. Registering the same weather tool through MCP produces the same tools entry, field for field, according to the dev.to post ^[5].

So MCP moves the plumbing behind that boundary rather than the boundary itself. The tool list now arrives from a process you keep running, and every tool the server exposes contributes its schema up front, whether this conversation needs it or not ^[7].

Treat "large context overhead" as a claim about somebody else's tool list. For it to hurt, the server has to expose many tools with long descriptions, and your sessions have to run long enough, or your window sit tight enough, that a fixed prefix on every request matters. The post publishes no token counts and no CPU or memory measurements ^[10], so the size of that tax is yours to measure before you argue about it.

The operational column is the part I would weigh at review: restarting a crashed server, sequencing startup order across several of them, and maintaining internal state over time ^[8]. A direct function call has none of those failure modes, though it inherits whatever your own process does when it falls over. If you own both sides and there is one client, the unified socket has one thing plugged into it.

What buys the overhead back is the case the post also names. Before MCP, every AI application wrote its own integration code for every external service it wanted to reach ^[9], and the post puts MCP's biggest strength in unifying that interface across a project's resources ^[11]. That is arithmetic about distribution. Ten applications against ten services is a hundred integrations the old way and twenty against one shared protocol ^[12]. One agent against three services you already own is three functions.

The CLI shortcut carries a different bill. Expose exactly one bash tool and you skip describing what bash can do, because the model has already read enough to write commands on its own ^[13]. The post is blunt that permission control here is hard to get right ^[14]. You have swapped schema maintenance for an authorization problem, which is the trade I would want stated out loud in a design doc rather than discovered in an incident.