When implementing an AI feature in a product recently, I noticed a tendency of management to steer towards a limited, well-behaved feature set that straight jackets the underlying model. This resulted in similar experience to what you describe. Maybe this is control and accountability thing? If I were to do this, I'd just slap a bunch of product specific tools (MCP, CLIs, HTTP API wrappers, etc) and skills (how to use those, best practices) with an agent and call it a day - if it can do more but also can fail, that's fine by me. That's why I like the idea of WebMCP more than custom built, limited AI chat interfaces that pop up everywhere nowadays. Just let Claude access everything and dump knowledge into it.