Claude Code Token Waste in Large Projects

Repowise introduces a local codebase index that pre-analyzes dependencies, git history, and architecture, providing MCP tools for Claude Code, Codex, and Cursor. In tests on the Django repo, agent tool calls dropped from 7.2 to 3.8, output reduced 31.6%. The tool supports 26 languages and runs without an API key.
Key insights
Problem: Token Waste in Large Projects
When using Claude Code on large projects, each new task requires re-reading files and understanding the code structure, wasting tokens on previously learned context. This inefficiency is especially pronounced in agent workflows where context is not preserved across tasks.
Repowise Solution: Local Codebase Index
Repowise builds a local index of the codebase, including code dependencies, git history, tests, documentation, and architecture decisions. It provides 10 task-specific MCP tools for Claude Code, Codex, and Cursor. Agents can query 'what files will be affected by this change?' and get cited answers without re-scanning.
Performance Results on Django Repo
In a test on the Django repository with 43 questions, agent tool calls decreased from 7.2 to 3.8 per task, and output was reduced by 31.6%. The methodology, sample, and lost lines are publicly available in the repository.
Additional Features: Risk Scoring and Code Health
Repowise computes a risk score (0-10) before code changes, based on recent commit distribution, and lists affected callers, relevant tests, and files that are often changed together. It includes 51 code health detectors covering defect risk, maintainability, and performance, providing refactoring suggestions that can be directly fed to agents.