AI Model Scanner

A small CLI that finds AI model files on your machine—Ollama, LM Studio, ComfyUI, Hugging Face cache, stray .gguf and .safetensors—and shows you what you have, where it lives, and which files are duplicates. Everything runs locally; no servers, no accounts.

---

Install

pip install ai-model-scanner

You need Python 3.9+. For faster scanning, install fd (macOS: brew install fd, Linux: apt install fd-find or dnf install fd, Windows: releases).

---

Quick start

ai-model-scanner scan

That scans your home directory for model files over 500MB and prints a table. To scan somewhere else or change the size cutoff:

ai-model-scanner scan --root ~/Documents
ai-model-scanner scan --min-size 100MB

Export to a file:

ai-model-scanner scan --export json > my_models.json
ai-model-scanner scan --export csv > my_models.csv

To see what would be scanned without doing it: ai-model-scanner scan --dry-run.

---

What it does

• Scan – Finds model files (.gguf, .safetensors, .pth, .bin, etc.) under given roots. Uses fd if present, otherwise find. Can stick to “known” tool paths (Ollama, LM Studio, ComfyUI, Hugging Face, MLX) or do a full tree.
• Incremental scans – After the first run it skips directories that haven’t changed, so later scans are faster. Use --no-incremental to force a full rescan.
• Duplicates – Hashes files (SHA256; for big files it samples the first 1MB) and groups duplicates so you can see wasted space.
• Code references – Can search your code (e.g. ~/Documents, ~/Projects) for references to model filenames. Handy before deleting duplicates.
• Cleanup / keep – cleanup finds unreferenced duplicate copies and can delete them (with confirmation). keep /path/to/model.gguf keeps that file and deletes the other copies with the same hash.
• Reports – Writes CSV/JSON reports of “referenced” vs “unreferenced” models (e.g. to Desktop).
• Watch – Watches paths and can notify when new large model files appear.

Results of a scan are cached so duplicates, cleanup, export, report, and show can reuse them without rescanning (see FAQ for cache details).

---

Commands (overview)

Command	What it does
scan	Find model files, optionally export, run health/reference checks
duplicates	List duplicate groups and (if requested) where they’re referenced in code
cleanup	Find unreferenced duplicates and optionally delete them
keep <path>	Keep one copy, delete the rest of its duplicate set
report	Write referenced/unreferenced model reports (CSV/JSON)
show	Show last scan results from cache
export	Export cached results to JSON/CSV/TXT
watch	Watch paths for new model files
health	Compare scan with Ollama/LM Studio CLI

Run ai-model-scanner <command> --help for options.

---

Configuration

Config file (create if you want to customize):

• macOS/Linux: ~/.config/ai-model-scanner/config.toml
• Windows: %APPDATA%\ai-model-scanner\config.toml

Example (see examples/config.toml.example for a full one):

[scanner]
min_size_mb = 500
known_paths_only = false
scan_roots = ["~/"]

[tools]
ollama_paths = ["~/ollama/models", "/usr/local/var/ollama/models"]
lm_studio_paths = [
    "~/Library/Application Support/LMStudio/models",
    "~/.lmstudio/models"
]

[output]
default_format = "table"
group_by_tool = true
show_duplicates = true

You can add your own paths under [tools] and set scan_roots for full scans.

---

Where it looks (defaults)

It knows about common installs for:

• Ollama – e.g. ~/ollama/models, ~/.ollama/models, platform-specific paths
• LM Studio – Application Support paths and ~/.lmstudio/models
• ComfyUI – ~/ComfyUI/models/checkpoints, loras, vae, etc.
• Hugging Face – ~/.cache/huggingface/hub
• MLX – macOS ~/mlx-community, ~/Library/Application Support/mlx

Exact paths are in the example config. If your models live elsewhere, add them to config.toml or run with --full-scan and optionally --learn-paths to have the scanner suggest new paths from what it finds.

---

Scan options (short reference)

• --root, -r – Directory to scan (default: home).
• --min-size, -s – Minimum file size (default 500MB).
• --full-scan – Ignore “known paths” and scan from roots (and disable incremental for that run).
• --no-incremental – Don’t skip unchanged dirs; rescan everything.
• --export, -e – json, csv, or txt.
• --dry-run – Print what would be scanned, don’t scan.
• --find-references – Search code dirs for model references.
• --health – Cross-check with Ollama/LM Studio.
• --learn-paths – Suggest new tool paths from discovered models and save to config.

---

Duplicates, cleanup, keep

• duplicates – Reads from last scan cache (or runs a scan). Shows groups of identical files and, if you asked for it, code references. Use --no-cache to force a fresh scan.
• cleanup – Same data, but finds unreferenced duplicate copies and asks before deleting. Use --dry-run to see what would be removed.
• keep /path/to/model.gguf – Keeps that file and deletes the other copies with the same hash (after confirmation). Use --dry-run to preview.

Cache: Commands that use the last scan (duplicates, cleanup, export, report, show, keep) use the cache by default. Run scan first if you haven’t, or pass --no-cache to force a new scan. There is no --no-cache on scan itself; use --no-incremental for a full rescan.

---

If something goes wrong

• “Command not found” – Ensure the environment where you ran pip install ai-model-scanner is on your PATH, or run python3 -m ai_model_scanner.cli scan.
• No models found – Check size threshold (--min-size 100MB), that paths are readable, and try --dry-run to see what’s included. If models are in a custom location, add it in config or use --full-scan.
• Permission errors – The tool skips files it can’t read. On macOS you may need Full Disk Access for some locations.

---

Development

git clone https://github.com/MichaelWeed/ai-model-scanner.git
cd ai-model-scanner
python3 -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -e ".[dev]"
pytest tests/ -v

Lint: ruff check ai_model_scanner/ tests/

Package layout: ai_model_scanner/ holds the CLI (cli.py), scanner, model/tool/duplicate logic, formatters, config, watcher, and reference finder. See pyproject.toml for deps and script entry point.

Contributions: open an issue or PR. Keep style consistent (PEP 8, ruff), add tests for new behavior, and update docs if you change behavior.

---

License

MIT. See LICENSE.