superclean

An agentic-dev garbage collector. One command, a tiered cleanup ladder, that reclaims the RAM, VRAM, and disk left behind by heavy parallel development, and never touches the editors, terminals, or AI tools you have open.

Quick start

uvx superclean-cli            # safe read-only report, changes nothing
uvx superclean-cli clean      # guided cleanup: confirm each proposed group
uvx superclean-cli sweep      # reclaim RAM/VRAM, kill orphaned dev servers

Or install it once (the command becomes superclean):

pipx install superclean-cli   # or: pip install superclean-cli

Windows, macOS, and Linux. Python 3.9+, one dependency (psutil). Start with the no-arg report, preview any tier with --dry-run, then run it. Full install and options below.

Why this exists

If you build with Claude Code (or any agentic workflow), your machine looks like this: several IDE windows, a wall of terminals, a fleet of dev servers, repeated Playwright runs, and one or more local models pinned in VRAM. Most of that never cleans up after itself.

Dev servers get orphaned when their parent shell dies. Playwright leaves stale browser builds. pip, npm, uv, and pnpm caches balloon. Models stay loaded long after you stopped using them. After a few days the system runs at capacity on garbage.

superclean is the collector for that garbage. It tells the difference between work you are actively doing and artifacts you are done with, and only removes the second kind.

The tiered ladder

Cleanup escalates through five additive tiers. Each includes everything lighter. A tier name describes intensity and risk, not a specific action, so the ladder is identical on every OS; each tier does what is available on the current platform and reports what it skipped.

Risk rises with the climb: tiers 1-3 are everyday-safe, wipe confirms, nuke makes you type the word. Prefer not to pick a tier? superclean clean runs a guided cleanup: it diagnoses the machine, proposes each action group with measured sizes (orphans, idle models, caches, old temp, targets.conf), and runs only what you confirm. Utilities: superclean report, superclean ram (RAM/VRAM relief only, no disk), superclean protected (the shield list), superclean init (scaffold the config files), and superclean last (replay the previous run).

The report also shows listening TCP ports with their owning process (ports held by orphaned dev servers are flagged) and GPU/VRAM usage (NVIDIA via nvidia-smi, AMD via sysfs). On macOS, listing other users' ports requires root; the section degrades gracefully.

The safety promise

superclean never kills your live tools. A generous baseline of editors, terminals, shells, and AI tools is protected by name, together with every one of their child processes, the entire ancestor chain of the running session, and any process whose command line shows it belongs to an AI agent. The tool also protects its own interpreter, so it can never flag itself. When it is unsure about a process, it leaves it alone. You add your own names in protect.conf.

Platform support

The everyday tiers do real work everywhere: on macOS and Linux, dust, sweep, and scrub genuinely kill orphaned dev processes, purge package caches, unload idle models, and age out temp files. Only the heaviest, most destructive actions differ by OS: on Windows they are handled by a proven PowerShell deep-clean backend that ships with the package; on macOS and Linux that native destructive deep-cleaning (browser/page cache, docker prune) is report-only in v1 (see Roadmap).

Install

The exact same commands work on Windows, macOS, and Linux (Python 3.9+).

The easiest path is uv (zero-install, always latest). If you do not already have uv:

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Then, on any OS:

uvx superclean-cli              # safe read-only report, changes nothing
uvx superclean-cli clean        # guided cleanup: confirm each proposed group
uvx superclean-cli sweep        # reclaim live resources

Prefer a permanent install (gives you the short superclean command)?

pipx install superclean-cli     # macOS: brew install pipx | Debian/Ubuntu: sudo apt install pipx
# or, simplest:
pip install --user superclean-cli

The only dependency is psutil (prebuilt wheels for Windows, macOS, and Linux, so there is nothing to compile). The package is superclean-cli on PyPI (the name superclean was taken); the installed command is superclean.

Recommended workflow

superclean                  # see what is going on, change nothing
superclean clean            # guided: confirm each proposed action group
superclean sweep --dry-run  # or preview a specific tier, then run it

Start at the lightest tier that solves your problem: sweep for "too many orphan processes and VRAM is full", scrub for "disk is filling up". Reach for wipe and nuke deliberately.

Configuration

Three optional files, shared by every platform (superclean init copies the commented examples into your per-user config dir). superclean looks for each file via SUPERCLEAN_CONF_DIR, then your per-user config dir, then the bundled examples. Lines starting with # are comments.

• protect.conf - extra process names to never touch (one per line).
• targets.conf - extra folders to age out at scrub (path|days|label). This is where machine-specific cleanup lives so the core stays generic.
• services.conf - extra local services to health-check in the report (label|url). Ollama is always checked.

Each ships with commented examples, including the author's own "Albert mode" setup, to copy from.

Scripting

Every command accepts --json for a stable machine-readable result (and suppresses human output), so superclean drops cleanly into pre-commit hooks, CI, or a scheduled idle-clean:

superclean report --json
superclean sweep --dry-run --json

Global flags: --dry-run, --yes/-y, --i-know (only with nuke), --quiet/-q, --json, --no-color, --log <path>, --force-unlock. Exit codes: 0 ok, 1 lock busy, 2 usage, 3 fatal.

Logs

Every run is logged in full under your per-user data directory:

• Windows: %LOCALAPPDATA%\superclean\
• macOS: ~/Library/Application Support/superclean/
• Linux: $XDG_STATE_HOME/superclean/ (or ~/.local/state/superclean/)

superclean last replays the newest mutating run from these logs.

Safety notes

superclean deletes files and stops processes. It is built to be conservative, but you are responsible for what you run on your machine.

• Use the no-arg report and --dry-run first. Always.
• wipe clears browser caches only when the browser is closed.
• nuke is destructive and irreversible on Windows, where it requires typing NUKE by hand unless you explicitly pass --yes --i-know. On macOS/Linux its destructive layer is report-only in this version.
• Provided as-is, no warranty. See LICENSE.

Roadmap

• Native destructive deep-clean for macOS and Linux (page cache, ~/Library/Caches, docker prune).
• A scheduled "garbage-collect on idle" mode.
• Per-project orphan attribution in the report.

Development

The portable core is Python (src/superclean/); the Windows deep-clean backend is PowerShell (windows/). See CONTRIBUTING.md.

License

MIT. See LICENSE.