from my understanding this is the high level differences
MCP apps live inside the tool itself so any AI client can render them (Claude, ChatGPT, etc), and once launched the app calls other tools directly without going back to the AI. Great for reviewing/approving data that already exists. also part of the protocol so not tied to a vendor.
Artifacts stay in conversation with the AI the whole time, so they're the better fit when you want to keep revising content back and forth. but i believe tied to claude.
Quick question for you, once you see an email draft, are you usually asking the AI to tweak or rewrite it, or mostly just reviewing and sending as is? if your looking to keep conversing with the LLM then and atifact is likely better. if its more review, make a change, and then approve/reject MCP app is great. The MCP app will also likely have lower token costs since its not involving the model every step