Call this at the START of a task. Reads the user's current Cursor usage/spend and returns conserve + followup flags. If conserve is true, ask any real questions through the questions/options UI (free) instead of open-ended prompts or silent defaults, and batch them into one prompt. If followup is true, end each task with a brief 'anything else?' options question. Auth is automatic (reads Cursor's local token); if needsLogin is true, ask the user to ensure Cursor is open/logged in on this machine, or to run the 'login' tool. Once the quota is exhausted the reading is cached until the cycle resets (pass refresh=true to force a fresh read).

FALLBACK auth only. By default get_usage needs no login — it reads Cursor's local token from state.vscdb. Use this tool only when that can't work: the MCP runs on a different machine than Cursor, the local DB is unreadable, node:sqlite is unavailable, or CURSOR_USAGE_FORCE_LOGIN is set. It opens a real browser, you log in, and it auto-discovers the usage endpoints and stores your session cookie locally (~/.cursor-usage).

Shows this billing cycle's usage broken down by model: cost, request count, and token totals (input/output/cache). Use when the user asks what's costing them or which models they use most. Heavier than get_usage, so call it on request rather than every task.

Clears the stored session cookie and discovered endpoints so get_usage reports needsLogin. Set forgetBrowser=true to also wipe the saved browser profile (forces a full re-login next time). Threshold config is preserved.

Set the minimum usage percentage (0-100) at which conserve mode activates. Default 0 means conserve whenever a reading succeeds and requests remain. Set e.g. 80 to only conserve near the limit.

Persists verbose mode: when on, get_usage returns a footer the agent appends to every message (Cursor Usage: X/limit requests · $used/$limit). Overridden by the CURSOR_USAGE_VERBOSE env var if set.

Persists follow-up mode (default off): when on, get_usage tells the agent to end each task with a brief 'anything else?' question via the questions/options UI, so you almost always get a prompt you can respond to. Overridden by the CURSOR_USAGE_FOLLOWUP env var if set.

Call this when the user declines an available update. Records the current remote version as skipped so get_usage won't surface it again — until an even newer version appears on GitHub.

Forces an immediate check against GitHub (bypasses the once/day throttle) and reports whether a newer version is available plus how to update. Normally the check runs automatically once a day.

Reports auth source, stored session/endpoints, thresholds, verbose/follow-up settings, and usage-cache age/cycle.