Donald Agent Hygiene
sterlingdon/donald-agent-hygiene
Audita e limpa skills/MCPs não usados em Claude Code e Codex com segurança
Ver no GitHubMCP server que audita e limpa skills, MCP servers, hooks e memória em Claude Code e Codex, identificando ferramentas não utilizadas que consomem contexto e degradam performance. Gera relatório HTML interativo, oferece dry-run seguro com backup automático e servidor web para limpeza com um clique.
Instalação
npx convez-ai sterlingdon/donald-agent-hygiene1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Remover MCP servers e skills instalados mas nunca utilizados
- Reduzir tamanho da janela de contexto e acelerar startup
- Auditar uso real de ferramentas através de logs de sessão
- Limpar duplicatas e clusters sobrepostos de funcionalidades
- Identificar arquivos de memória oversized e hooks obsoletos
Veredito

Ferramenta essencial para manutenção de ambientes Claude Code e Codex em produção. Design orientado à segurança (read-only scan, dry-run, backups) e transparência (relatório detalhado) torna viável remover cruft acumulado sem risco — impacto direto em performance e custo de tokens.
README
A tradução para o português deste README ainda está na fila.
donald-agent-hygiene 🧹
Keep Claude Code and Codex in their healthiest state — audit every place they keep skills / MCP servers / hooks / memory (including plugin-bundled ones), cross-reference your real usage from session transcripts, and safely clean out the cruft you installed but never use.
We accumulate MCP servers and skills without noticing. Many are installed-but-unused — they bloat the context window, slow startup, and degrade tool selection. This tool finds them and helps you remove them, across both Claude Code and Codex, in one pass.
stdlib-only Python · no third-party runtime deps · 64 tests · scan is always read-only.
What it does
- Inventory every source: personal / project / plugin-bundled skills, user / project / plugin MCP servers, hooks, and memory files — for Claude Code and Codex.
- Usage from
~/.claude/projects/**/*.jsonland~/.codex/sessions/**— counts how often each skill / MCP was actually invoked, and when last. - Classify each item 🟢 safe-to-clean / 🟡 review / 🔴 caution / ✅ keep — flags never-used, stale, duplicate, overlapping clusters, high-cost-low-use MCP, disabled-plugins-on-disk, oversized memory files.
- Report as a self-contained interactive HTML page.
- Clean safely:
apply(dry-run by default,--applyexecutes with automatic backup) or a local one-click-execute web server.
Usage
cd scripts
# Read-only audit -> interactive HTML report
python -m hygiene.cli scan --home "$HOME" --codex-home "$HOME/.codex" \
--projects "$HOME/.claude/projects" --sessions "$HOME/.codex/sessions" \
--cwd "$PWD" --out /tmp/report.html
# Preview cleanup (dry-run). Add --apply to execute (backs up first).
python -m hygiene.cli apply --home "$HOME" --codex-home "$HOME/.codex" \
--projects "$HOME/.claude/projects" --sessions "$HOME/.codex/sessions" \
--cwd "$PWD" --severity green
# Local server: click "execute" per item in the browser
python -m hygiene.cli serve --home "$HOME" --codex-home "$HOME/.codex" \
--projects "$HOME/.claude/projects" --sessions "$HOME/.codex/sessions" \
--cwd "$PWD" --port 8765
python -m pytest -q # run from scripts/
Safety
- Scan is always read-only. Cleanup defaults to dry-run; mutations require
--apply(CLI) or a click (server). - Every mutation backs up first to
~/.agent-hygiene-backups/. - The local server binds 127.0.0.1 only, gates execution behind a random token, and asks for confirmation.
- Codex
config.tomlis never auto-edited (MCP removal is emitted as acodex mcp removecommand). - Hooks and memory are never auto-cleaned; Codex skills (whose usage signal is best-effort) are never auto-green.
Layout
scripts/hygiene/ collect_cc, collect_codex, usage_cc, usage_codex, cost,
classify, report, actions, serve, cli, model, normalize, util, paths
scripts/tests/ pytest suite
docs/ research notes, design, and the M1/M2/M3 implementation plans
SKILL.md agent-skill manifest (Claude Code / Codex)
License
MIT