Reduce Claude Code token usage
If you're hitting your Claude Code usage limit mid-session, the problem usually isn't your prompts — it's the tool output. File reads, grep results, test logs, and command output pile up in your context window fast. TokenPincher compresses that output before it reaches the model, lossless by default.
npx tokenpincher initclaude mcp add tokenpincher -- npx -y tokenpincher mcpThe usage-limit problem
Claude Code is agentic — it reads files, runs commands, greps across your codebase, and reads the results back into context. That tool output is usually far larger than anything you typed. On a long session it's often the majority of your token spend, which is why usage limits hit hardest on exactly the sessions where you're being most productive.
How TokenPincher compresses it
TokenPinchersits between Claude Code and the model as a thin CLI wrapper (or an MCP server, if you'd rather configure it that way). It shrinks large tool outputs and logs, lossless by default — the model still gets what it needs, just without the redundant bulk. Outputs over 400 lines are truncated head and tail with a visible "N lines omitted" marker. Nothing about how you use Claude Code changes.
No desktop app
One line, no menu-bar app, no background process to babysit. It runs invisibly alongside Claude Code.
Also using Codex?
TokenPincher works the same way for OpenAI Codex CLI. See reduce Codex token usage.
For the deep dive — where the tokens actually go and how compression works — read the full guide.
Get started
npx tokenpincher init