Home / Blog / Guides

Guides

How to cut token usage in Claude Code, Cursor and Codex

Coding agents burn most of their tokens reading, not writing. The fastest way to cut cost is to stop sending whole files: retrieve code by meaning, return exact line ranges, and let the agent ask for callers or tests only when it needs them.

Published October 2, 2026 · 6 min read

Where the tokens actually go

Watch a coding agent work on an unfamiliar repo and the pattern is always the same: grep for a keyword, open the top five matches, skim them, open three more, then finally edit ten lines. Every file it opened went into the context window in full. A 600-line file costs roughly 6–8k tokens whether the agent needed 600 lines or 15.

On a typical "where is this handled?" question we measured about 11,480 tokens for the read-the-files approach versus 214 tokens for the snippet Finder returned — about 98% less for the same answer.

Five ways to shrink context

1. Retrieve by meaning, not by keyword

Grep finds text. An agent asked to fix token refresh finds every file that says "token" first. Semantic search finds TokenManager.refresh() because it matches intent. Fewer wrong files opened means fewer tokens spent.

2. Return line ranges, not files

The unit of context should be the snippet: path#L142-178 with numbered lines. Finder's file-finder tool does exactly this when the agent already knows the file but not the range.

3. Use the call graph instead of re-reading

"Who calls this?" should be a graph lookup, not a repo-wide grep followed by five file reads. Finder's context tool answers it from a local graph — no embedding call, no credits.

4. Check impact before editing

A failed edit is the most expensive tokens of all: the agent reads, edits, runs tests, reads the failure, reads more files. impact lists the files and tests a change touches up front.

5. Teach the agent the workflow

Tools only help if the agent uses them in the right order. Finder's one-click setup writes prompt and skill files (finder-kit) that route each question to the narrowest tool: broad discovery once, then exact lookups.

Setting it up

Install Finder, open the dashboard, and click your harness — Claude Code, Cursor, Codex, VS Code, Gemini CLI, Windsurf and nine others are supported. The first query on a workspace asks you to approve that exact folder before anything is indexed.

See the setup guide for per-harness details.

Frequently asked questions