MCP tokens get scopes — say exactly what an agent may do, and see what it did
An MCP token used to be all-or-nothing: connect Claude Desktop or any other agent and it could do everything your connection allows, with no way to hand one client read-only access. Every token now carries tier checkboxes — Read (list notes, account status, best times), Draft (save drafts, library writes, repurpose), Publish (schedule real publishing) — and every call is checked against those tiers at a single gate before anything runs. A token without Publish checked can never schedule real publishing, no matter how it's prompted.
- New tokens default to Read + Draft; Publish has to be ticked on purpose. Tokens issued before this change keep working untouched and pick up full scope on their next call.
- A new usage panel sits under each token in Settings → AI → MCP & API: calls by day over the trailing week with failures counted, plus the last 50 requests with latency — denied attempts included, so a mis-scoped client shows up immediately.
- Each token lists its actual access as chips right in the token list, so there is never a guess about what a connected agent can do.
- Scopes only ever narrow what your account settings already allow — switching off "Allow agents to schedule/publish" still wins regardless of how a token was issued.
Nothing to reconfigure: existing integrations behave exactly as before, and revoking a token deletes its usage history along with it.