Token limits

Reduce Claude Code token usage

Practical ways to spend fewer tokens in Claude Code sessions, including local memory, scoped reads, and workflow discipline.

Why tokens disappear fast

Claude Code sessions grow expensive when the model re-reads large files, re-explores the repo each turn, or carries long chat history. The fix is not a single setting: it is giving the agent a map and a process that avoids redundant context.

Use a local codebase map first

AgenKit Local Memory builds a deterministic index on your machine. Agents query file:line pointers instead of opening whole trees. Phase 0 Orient refreshes that map before /agenkit runs.

  • Run npx agenkit memory status when answers feel stale
  • Let /agenkit Orient run before large tasks
  • Prefer memory query over broad grep in chat

Keep tasks scoped

Split work into approval-gated phases (spec, architecture, plan, build). Smaller prompts with clear file paths beat one giant instruction that forces the model to search.

Related docs → · AgenKit for claude-code → · Pricing →