aiclean

aiclean

Scan and safely clean local caches left behind by AI coding tools — Claude Code, Cursor, VSCode, Ollama, npm/pnpm/yarn, uv/pip, HuggingFace, PyTorch, Playwright, Hermes, and more.

AI dev tools are notorious for quietly eating disk space: Ollama models, Claude Code project logs, HuggingFace model caches, browser binaries for test automation... it adds up to tens of gigabytes without anyone noticing. aiclean finds it, shows you exactly where it is, and only deletes what you explicitly confirm.

Install

pip install aiclean
# or run without installing:
pipx run aiclean

Usage

# Scan only (read-only, safe to run anytime)
aiclean

# See all known cache locations this tool checks
aiclean --list

# Preview what would be deleted (no changes made)
aiclean --clean --dry-run

# Interactively confirm each cache before deleting
aiclean --clean

# Clean specific targets without prompts (careful!)
aiclean --clean --only ollama-models npm-cache --yes

Example output

      SIZE  NAME                     PATH
----------------------------------------------------------------------
    12.4GB  ollama-models            /Users/you/.ollama/models
     3.1GB  claude-code-projects     /Users/you/.claude/projects
     1.8GB  huggingface-cache        /Users/you/.cache/huggingface
   420.0MB  npm-cache                /Users/you/.npm
   180.2MB  uv-cache                 /Users/you/.cache/uv
----------------------------------------------------------------------
    17.9GB  TOTAL (5 caches found)

Run with --clean to interactively remove caches (add --dry-run to preview safely).

Safety

• Scan mode never touches the filesystem — it only computes sizes.
• --clean without --yes asks per-target before deleting anything.
• --dry-run combined with --clean shows exactly what would be removed, with zero deletions.
• Targets are well-known, documented cache directories only — aiclean never does a generic recursive delete of arbitrary paths.

What it scans for

Run aiclean --list for the full, current list. Broadly: editor/IDE caches (Cursor, VSCode), AI coding agent caches (Claude Code, Hermes), local model caches (Ollama, HuggingFace, PyTorch Hub), and package manager caches (npm, pnpm, yarn, uv, pip) — since these are the biggest and most overlooked space users of an AI-assisted dev setup.

Contributing

Issues and PRs welcome — especially "please add support for X cache location" requests, since that's exactly the kind of real-world signal this project is looking for.

License

MIT