Repowise claims 96% fewer tokens for Claude Code context-loading, but its own README tells a more modest story

A repo-indexing tool called Repowise is getting attention for supposedly cutting Claude’s token use by 96%, but the numbers in its own README tell a quieter story. The project has real traction (5,000+ stars, 540 forks), and the core idea is sound. Instead of making Claude grep through your entire codebase every time it needs context, Repowise indexes your repository locally and serves up 10 task-specific tools. Ask for a function definition, get a function definition. Ask for dependencies, get dependencies. No raw file dumps.

The disconnect is in the marketing. The promotional post claimed 96% fewer tokens and 70% fewer tool calls. The actual benchmark in the README, run against the Django repo, shows 31.6% fewer output tokens and a drop from 7.2 to 3.8 tool calls per task. That’s still useful, but it’s not the order-of-magnitude leap the headline suggests. The difference matters if you’re paying per token or trying to stay under context limits.

What it means for your workflow

Repowise competes with tools like CodeGraph, and the pattern is worth understanding. If your team uses Claude for code review, refactoring, or documentation, a local index can cut noise and cost. The 10 tools cover common tasks like finding definitions, tracing imports, and listing files by type. It runs as an MCP server, so it plugs into Claude Desktop or any MCP-compatible client.

The lesson is to test the actual numbers in your environment. A 30% token reduction is real money at scale, and fewer tool calls mean faster responses. This is exactly the kind of efficiency layer we build into an AI command centre, where context routing and tool selection happen behind the scenes. Just don’t expect 96% savings until you’ve run your own benchmark.