OpenWolf cuts Claude Code’s token burn with a project-wide cheat sheet

What it does

OpenWolf injects six Node.js lifecycle hooks into Claude Code. Before Claude opens a file, the hook serves up a project map (anatomy.md) with file descriptions and token estimates; if Claude already read the file this session, OpenWolf warns it. After edits, it auto-updates the map and appends to a chronological memory log. It also maintains cerebrum.md—a running list of your corrections and preferences—and buglog.json, so Claude can search past fixes instead of rediscovering them. The authors report an average token reduction of 65.8% across 20 projects and 132+ sessions, with about 71% of repeated file reads caught.

The interesting bit

The whole system is just a .wolf/ directory of markdown and JSON files, plus hooks. There is no separate AI runtime; it simply feeds context to Claude and relies on Claude following instructions to update its own “brain.” The token ledger even tracks anatomy cache hits and misses, making the tool accountable to its own savings claims.

Key highlights

• Claims ~80% token savings versus bare Claude CLI on a large active project, with estimates derived from a character-to-token ratio.
• cerebrum.md acts as a Do-Not-Repeat list: past mistakes, style preferences, and project conventions Claude is supposed to consult before writing code.
• buglog.json stores root-cause/fix pairs so Claude can search known solutions before debugging.
• Includes side utilities: a designqc screenshot tool (via Puppeteer) for UI evaluation and a reframe command that curates migration prompts for 12 UI frameworks.
• Falls back to CLAUDE.md instructions if Claude Code hooks fail to fire.

Caveats

• Token counts are estimates, not exact API tallies; the README admits a ~15% margin of error.
• The learning memory (cerebrum.md) relies on Claude voluntarily updating it, and compliance is only ~85–90%.
• Claude Code hooks are a new feature; if they misfire, OpenWolf degrades to static instructions.
• The project is at v1.0.4 and the authors note that things may break.

Verdict

Worth a look if you’re burning through context windows with Claude Code on large or long-lived codebases. Skip it if you want guaranteed, exact token accounting or a fully deterministic agent memory.