Claude Code token-saving plugins

Core ideas

• Headroom compresses context: It sits between you and your AI model, stripping unnecessary lines so you get the same output using far fewer tokens.
• Reversible compression: The fact-checker adds that Headroom's compression isn't lossy. Claude can pull the original text back anytime, and it has almost 30,000 GitHub stars.
• Graphify builds a knowledge graph: It maps your whole codebase, linking files and functions so Claude doesn't need to reread everything on every message.
• Karpathy-inspired tool: Graphify is credited to ideas from Andrej Karpathy, and on large projects it can cut token costs up to 70 times, a number the narrator says was actually undersold.
• Code Burn tracks spend: It breaks down token burn by cost, task, model, tool, and project, runs fully local with no API keys, and your code never leaves your laptop.
• Ponytail's real number gets corrected: The original claim was over 50% usage reduction with no accuracy loss. The fact-check says independent tests actually show 54% less code but only 22% fewer tokens, still solid but not the advertised 50%.
• Structure of the video is a claim-and-correct format: One voice makes the pitch, a second voice verifies or challenges each stat live, which is the hook of the piece.

Quotes