Session 40, checked via /context: this project’s own .mcp.json only declares 4 servers (txtavern, pixellab, txtavern-postgres, honeybadger). Everything else present in the session — WIN, ngk, Gmail, Google Calendar/Drive, Microsoft 365, flowbite, playwright — is a Claude.ai account-level connector attached regardless of project, contributing ~240k tokens (24% of the 1M context window) just in deferred-tool listing overhead, before any full schema is even loaded. None of those servers were called once during a 5-hour live D&D session.
This is a real, measured contributor to “the assistant feels slow” during live play — every turn’s prompt processing scales with total context, and a quarter of it is inert. Not fixable from inside a Claude Code session; needs either (a) disconnecting irrelevant connectors in Claude.ai account settings before a DM session, or (b) per-project connector scoping if that becomes available, so this project’s sessions only ever see txtavern + pixellab (+ postgres/honeybadger for maintenance work).
Related, same session: even without the connector bloat, a single long-running conversation that carries an entire live session end-to-end reaches hundreds of thousands of tokens in messages alone (694.9k / 69.5% of context in this session). Worth establishing a convention of resetting context (/compact or a fresh conversation) at a natural act break during a long session, not just at session-open.