Claude Config Auditor — Development skill for Claude Code

claude-config-auditor

[](https://pypi.org/project/claude-config-auditor/) [](https://pypi.org/project/claude-config-auditor/) [](https://github.com/emreyildirim/claude-config-auditor/actions/workflows/tests.yml) [](https://opensource.org/licenses/MIT)

A **non-destructive** linter and cost auditor for `.claude/` and `CLAUDE.md` — *we annotate, we don't rewrite*. Measures the **token cost** your Claude Code config pays on every session, and audits **agent / skill quality** (missing descriptions, overlapping routes, broken YAML).

Think of it as ESLint for your context window.

Why this exists

Claude Code loads `CLAUDE.md`, every `.claude/agents/*.md`, and every skill's `SKILL.md` into the context window on every session start. That's a **fixed per-session tax** — and most projects don't know how big theirs is.

The ecosystem has a lot of "session handoff" and "state management" tools. It doesn't have a linter for the config itself. This is that linter.

• "How many tokens does my CLAUDE.md actually cost?"
• "Are two of my agents describing the same job?" (Anthropic's docs say agents are selected based on their description; the exact ranking algorithm isn't published, but overlapping descriptions create routing ambiguity in practice.)
• "Is my SKILL.md description too vague for Claude to ever invoke it?"

Cost snapshot — five popular frameworks

The headline `claude-audit --html` numbers from a clean install of five popular Claude Code frameworks, audited with `--semantic` on:

Framework	Always loaded	Window	Files	Findings
BMAD