building internal agent tooling, when does an MCP server beat just giving the model a function schema
working on some internal agent stuff at work and im genuinely unsure where the line is. for a single app, defining tools inline as function schemas is dead simple and i control everything. MCP feels like it earns its keep the moment two different apps want the same tool, or when a non-dev on my team wants to plug notion in without touching my code. is that the right mental model or am i missing something. and does anyone route SAE-style introspection through a tool the agent can call, or is that a dumb idea
MCP wins the second more than one agent (or more than one human) needs the same tool. if its just your one script talking to your one model, a function schema is less moving parts and honestly less to get owned. i got handed 3 of these and the MCP ones are the only ones i can reason about the blast radius on. so: reuse across clients = server, one-off = schema