Loading page content
Audit what is actually filling your Claude Code context: instruction files, tool definitions, loaded documents, and find the tokens you are paying for on every single session without benefit.
Run /context in Claude Code, paste the output here, and get a waste breakdown with tier recommendations and the dollar cost of your startup overhead. Runs entirely in your browser; nothing is sent anywhere.
Thresholds are heuristics from real audits, not gospel. The method behind this tool: context engineering is FinOps for your AI agent. Related: MCP vs. skills, when to use subagents, and the CLAUDE.md generator.
Context is a recurring cost, not a one-time one. Everything loaded at session start is paid for on every request in that session, so a 3,000-token instruction file is not a 3,000-token cost. It is 3,000 tokens times every turn of every session, forever. Auditing it is one of the few optimizations that compounds.
The usual culprits are MCP server tool definitions that are always loaded and rarely called, instruction files that grew by accretion because nobody deletes, and documentation loaded wholesale when a pointer would do. The test for any line is whether you can name a mistake it has prevented. If not, it is idle spend.