Why

By default, Claude Code loads every enabled plugin's metadata into your context on every turn. For a fresh install with the default plugin set, that's often 500,000+ tokens of agent descriptions, skill metadata, and MCP preambles — regardless of whether you ever invoke them.

The result: shorter conversations before compaction, slower responses, higher cost, and a tax on multi-agent workflows. claude-lean makes that hidden cost visible — and gives you four levers to reclaim it.

The Four Levers

┌──────────────────────────────────────────────────────────────────┐
│                                                                   │
│   1. AUDIT     →   See exactly what's eating your context        │
│   2. APPLY     →   Disable what you don't use (with backup)      │
│   3. PROFILE   →   Different active set per project / cwd        │
│   4. RESTORE   →   One command back to any previous state         │
│                                                                   │
└──────────────────────────────────────────────────────────────────┘

Quick Start

# Install (zero hard dependencies beyond rich, which most setups have)
pipx install claude-lean

# 1. See what's costing you
claude-lean audit

# 2. Apply recommendations interactively (with diff + backup + dry-run safety)
claude-lean apply

# 3. Switch to a focused profile in a specific project
cd ~/my-python-project
claude-lean profile use python-ml

# 4. Undo anything
claude-lean restore --latest

Real Numbers From a Real Setup

A representative audit on a default-everything install:

─────────────────────────  claude-lean audit  ─────────────────────────

  Plugins enabled: 20  ·  Skills: 262  ·  Agents: 18
  Estimated system-prompt overhead per turn: 593,122 tokens

  Top contributors (tokens / turn):
    engineering-advanced-skills   129,302   ✓ keep
    engineering-skills            123,232   ✓ keep
    marketing-skills              111,126   ⚠ disable (no marketing work)
    c-level-skills                 68,660   ⚠ disable (no exec work)
    ra-qm-skills                   44,325   ⚠ disable (no regulatory work)

  17 findings · Estimated savings if applied: 298,283 tokens / turn
                                              ≈ 2.0× more headroom

With per-project profiles (level 3 above), that 2.0× compounds to 5×+ by loading only the plugins relevant to that project's stack.

How It Works

claude-lean is a standalone Python CLI, not a Claude Code plugin. It runs outside Claude Code and operates on the plain files in ~/.claude/. Because it never loads into your prompt context, it costs zero tokens per Claude conversation.

                ┌─────────────────────────────────┐
                │         claude-lean CLI          │
                │     (runs in your terminal)     │
                └────────────────┬────────────────┘
                                 │ reads / writes
                                 ▼
                ┌─────────────────────────────────┐
                │           ~/.claude/             │
                │  • CLAUDE.md                    │
                │  • settings.json                │
                │  • plugins/cache/               │
                │  • projects/*/memory/           │
                └─────────────────────────────────┘

See docs/how-it-works.md for the full architecture.

Safety

• Every write creates a backup first to ~/.claude/.claude-lean-backups/{timestamp}/
• Dry-run is the default for non-TTY usage — pipes don't accidentally mutate state
• claude-lean restore --latest reverts the most recent change in one command
• --list shows every snapshot you've ever made

See docs/safety.md for the full safety model.

Documentation

Commands

Status

v0.1.0 — alpha. The core (audit, apply, restore, profile) is working and tested. The following are roadmapped for later versions:

• v0.3 — claude-lean monitor (live token-usage dashboard from session logs)
• v0.4 — Claude Code hooks integration (auto-switch profile on cd)
• v1.0 — Profile marketplace + community-contributed profiles

License

MIT — see LICENSE.

Contributing

Issues and PRs welcome. Per-project profiles are intentionally easy to contribute — drop a TOML file in src/claude_lean/profile/stock/, add a test, ship a PR.