Tokenminning in Cursor

Cursor re-sends context on every agent step: rules, attached files, MCP tool schemas, and growing chat history. Most spend comes from long threads and heavy configuration—not one verbose reply.

Hub for the Cursor cluster — optimization, routing, and configuration. For spent vs remaining allowance, token breakdown, and billing controls, use the dedicated Cursor usage dashboard at cursor.com/dashboard. For the general stack see Where to start.

Quick checklist

1. Open the Cursor usage dashboard — spent vs remaining allowance at cursor.com/dashboard.
2. Use Auto or Composer (First-party pool) for routine agent work.
3. Shorten .cursor/rules and disable MCP you are not using.
4. New chat per task—not one marathon thread.
5. Audit Memories and user rules if input stays high.

Typical impact: 60–90% savings switching models; 20–50% trimming rules/MCP; 30–60% less cache growth from shorter chats.

How Cursor bills (overview)

Each agent turn sends your prompt plus attachments. Follow-ups reuse context as cheaper cache read tokens until you start a new chat.

Two monthly usage pools on individual plans: API (named frontier models) and First-party (Auto, Composer 2.5, Grok 4.5). Pool meters and billing tabs: Cursor usage dashboard.

Guides

Model routing

Model routing on Cursor:

Switching models mid-chat resets cache — start a new chat.

Trim configuration

Anti-patterns

Related

• IDEs
• Alternatives to Cursor
• Cursor docs