civitai-hub

huggingface_hub, but for CivitAI. Point it at a model URL to inspect what's inside (type, base/parent model, every file with sizes and hashes) and download the checkpoint or LoRA into a deduplicated, content-addressed cache — usable as both a CLI and an importable Python library.

$ civitai download https://civitai.com/models/580857 --fp16 -o ~/ComfyUI/models/loras
~/ComfyUI/models/loras/realistic-skin-xl.safetensors

civitai info <url> — what's inside a model at a glance

Contents: Features · Install · CLI · Library · Auth · Configuration · How it works · Docker · Development

Features

• Inspect before you fetch — civitai info <url> shows the model type, base/parent model, the selected version, and a table of every file in it (size, format, precision, hash, scan status).
• Smart, scriptable downloads — picks the URL-pinned version (or the latest) and its primary file by default; override with --version-id, --fp16/--fp32, --pruned/--full, --format, --file, or grab everything with --all.
• Find the base model — civitai base <lora-url> lists the checkpoints a LoRA's base-model family runs on (CivitAI only exposes the family name, not a direct link), with --download N to grab one.
• Discover & identify — civitai search finds models from the terminal; civitai by-hash <file> tells you what an orphaned .safetensors actually is.
• Manage it — civitai cache ls/verify/rm/prune for the local store, civitai login to save your token, --json on every command for scripting.
• Managed cache like HF Hub — content-addressed blobs keyed by SHA256, with per-version snapshot symlinks. The same file reused across versions is stored once. Re-downloads are skipped.
• Resumable & verified — HTTP range-resume for interrupted downloads, automatic SHA256 verification, and a live progress bar.
• Library + CLI — everything the CLI does is a one-line call from Python.
• Safe by default — blocks files flagged unsafe by CivitAI's scanners (override with --allow-unscanned) and fails fast on early-access/gated content with a clear message.

Install

From PyPI (recommended):

pipx install civitai-hub        # isolated CLI
pip install civitai-hub         # into the current environment

With Docker — no install needed:

docker run --rm ghcr.io/mr8bit/civit-ai-cli info 580857

From a GitHub release (without PyPI) — pin a tag from the releases page:

pipx install "git+https://github.com/mr8bit/civit-ai-cli@v0.3.0"

For development from a clone:

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

CLI quickstart

# Inspect a model (human table, or --json for machines)
civitai info https://civitai.com/models/580857/realistic-skin
civitai info 580857 --json

# Download the primary file of the latest (or URL-pinned) version
civitai download https://civitai.com/models/580857

# Drop a specific precision straight into your ComfyUI/A1111 folder
civitai download 580857 --fp16 -o ~/ComfyUI/models/loras

# Preview what would be fetched without downloading
civitai download 580857 --dry-run

# Grab every file in the version
civitai download 580857 --all

# Find the base checkpoints a LoRA runs on, then grab one
civitai base 580857
civitai base 580857 --download 1 -o ~/ComfyUI/models/checkpoints

# Search, then identify a mystery file on disk
civitai search "realistic" --type Checkpoint --base-model "SDXL 1.0" --limit 10
civitai by-hash ~/ComfyUI/models/loras/mystery.safetensors

# Save your token; manage the cache
civitai login                       # prompts for the API key (or --token)
civitai cache ls
civitai cache verify
civitai cache rm 580857

civitai base <url> — the checkpoints a LoRA runs on, most-downloaded first

The downloaded path is printed to stdout (the progress bar goes to stderr), so it pipes cleanly:

MODEL=$(civitai download 580857 --no-progress)

Library usage

import civitai_hub

# Inspect
info = civitai_hub.model_info("https://civitai.com/models/580857")
print(info.model.type, info.version.base_model, len(info.files))

# Download — returns the local Path (or list[Path] with all=True)
path = civitai_hub.download(
    "https://civitai.com/models/580857",
    fp="fp16",
    local_dir="~/ComfyUI/models/loras",
)

download(...) mirrors hf_hub_download (single file) and snapshot_download (all=True).

Authentication

Most public files download without a token. For gated, NSFW, or early-access resources, create an API key in your CivitAI account settings and provide it via:

export CIVITAI_TOKEN=<your key>     # or pass --token <key>

Configuration

Precedence is flag → environment variable → default.

The flags live on download/base; civitai info takes only --version-id/--json and reads the env vars for everything else.

Mirror (civitai.red). If civitai.com is blocked for you, paste a civitai.red/… URL and the API and downloads route through it automatically; for bare ids or search, set CIVITAI_HOST=civitai.red. Only civitai.com and civitai.red are trusted, so your token can't be redirected to some other host.

How it works

A model URL resolves to one GET /api/v1/models/{id} call; the version is the URL's pinned modelVersionId (or the latest published one), and the file is that version's primary .safetensors unless you filter it. Downloads go only to civitai.com (the host is checked first) and authenticate with the Authorization header — httpx strips it on the cross-host CDN redirect, so the token never leaves civitai.com or lands in a URL. The download is SHA256-verified and lands in the cache:

$CIVITAI_HOME/
└── models/<modelId>/
    ├── blobs/<sha256>                      # content-addressed, deduplicated
    └── snapshots/<versionId>/<filename>    # symlink → ../../blobs/<sha256>

--local-dir then materializes a symlink (or copy) of the blob into the folder you choose.

For the full command/flag reference, cache details, troubleshooting, and exit codes, see docs/usage.md.

Docker

Prebuilt multi-arch images are published to the GitHub Container Registry on every release:

# inspect a model
docker run --rm ghcr.io/mr8bit/civit-ai-cli info 580857

# download into a host folder (mount it as the cache + output)
docker run --rm -v "$PWD/models:/data" \
  ghcr.io/mr8bit/civit-ai-cli download 580857 -o /data

# pass a token for gated content
docker run --rm -e CIVITAI_TOKEN="$CIVITAI_TOKEN" \
  ghcr.io/mr8bit/civit-ai-cli base 580857

The cache lives at /data inside the image (CIVITAI_HOME) — mount a volume there to persist downloads. Build it yourself with docker build -t civitai-hub ..

Development

pip install -e ".[dev]"
pytest                 # full suite (offline; httpx mocked with respx)
pytest tests/test_resolver.py::test_fp_selector -v   # a single test
ruff check src tests   # lint
CIVITAI_LIVE=1 pytest tests/test_live.py -v           # opt-in test against the real API

See CONTRIBUTING.md for the architecture overview and conventions, and docs/superpowers/specs/ + docs/superpowers/plans/ for the original design and implementation plan.

License

MIT © Artemiy Mazaew