CLI

Every command, grouped the way the help output groups them.

hf <command> [arguments] [flags]

Run hf <command> --help for the full flag list on any command. Every command that takes a <ref> accepts a bare name, an owner-qualified id, a huggingface.co URL, or an hf:// URI.

Read one thing

Command	What it does
`get <ref>`	Read whatever a reference points at
`model <ref>`	One model, every field the hub returns
`dataset <ref>`	One dataset
`space <ref>`	One space, runtime state included
`kernel <ref>`	One compute kernel
`user <name>`	One user profile
`org <name>`	One organization
`ns <name>`	A name without knowing whether it is a user or an org
`collection <slug>`	One collection and its items
`paper <id>`	One paper, its authors, and what cites it
`post <ref>`	One community post
`blog <slug>`	One blog entry, body included
`discussion <repo> <num>`	One discussion or pull request, with its event thread
`whoami`	Who the current token belongs to

List

Command	Key flags
`models`	`--author`, `--search`, `--filter`, `--task`, `--library`, `--language`, `--license`, `--sort`, `--direction`, `--full`
`datasets`	the same, plus `--benchmark`, `--size`
`spaces`	the same, plus `--sdk`
`kernels`	`--author`, `--search`, `--sort`
`collections`	`--owner`, `--item`, `--sort`
`papers`	`--search`, `--date`, `--daily`, `--sort`
`posts`	community posts
`blogs`	from the RSS index
`discussions <repo>`	`--status`, `--type`, `--author`
`search <query>`	`--type`, across every entity kind
`trending`	`--type`

The filter flags are sugar over the one filter= parameter the API takes, joined with AND, so --task text-generation --library transformers and --filter text-generation --filter transformers are the same request.

--item on collections is the reverse lookup: given a repo, which collections contain it.

Command	What it does
`likers <repo>`	The users who liked a repository
`likes <name>`	What a user has liked
`followers <name>`	A user's followers
`following <name>`	Who a user follows
`members <name>`	An organization's members
`upvoters <ref>`	Who upvoted a paper or a collection

Repository contents

Command	What it does
`refs <repo>`	Branches, tags, and conversions
`tree <repo> [path]`	The files at a path, `-r` to recurse, `--commits` for per-entry detail
`files <repo>`	Every file, flat
`commits <repo>`	The commit log, `--rev` to pick a revision
`paths <repo> [path...]`	File metadata and security scans for named paths
`card <repo>`	The parsed README front matter
`readme <repo>`	The card, written to stdout
`cat <repo> <path>`	One file, written to stdout
`page <ref>`	A rendered page as JSON, 1:1 with what the browser gets

Datasets

The dataset viewer is a second API on its own host, and these commands read it.

Command	What it does
`valid <dataset>`	Which viewer capabilities a dataset has
`splits <dataset>`	Configs and splits
`size <dataset>`	Row and byte counts
`schema <dataset>`	Column types for a split
`stats <dataset>`	Per-column statistics
`rows <dataset>`	Rows, with `--offset`, `-n`, `--order-by`
`head <dataset>`	The first rows of a split
`dsearch <dataset> <query>`	Full text search within a split
`filter <dataset> <where>`	Filter a split with a SQL-ish predicate
`parquet <dataset>`	The converted parquet shards
`croissant <dataset>`	The hub's own Croissant metadata

Graph

Command	What it does
`uri <input>`	Resolve any reference to its canonical `hf://` URI
`url <input>`	Resolve any reference to its https location
`graph <ref>`	The node and its edges
`edges <ref>`	Edges only
`children <model>`	The models derived from a model
`parents <model>`	The models a model was derived from
`citations <paper>`	The repositories tagged with a paper
`crawl <ref>`	Walk breadth-first, `--depth`, `--follow`, `--max-nodes`, `--max-requests`, `--nodes-only`, `--edges-only`
`rdf <ref>`	Write one entity as RDF, `--format nt\|ttl\|jsonld\|nq`
`export <ref>`	Write a whole graph to one file, `--format jsonl\|json\|nt\|ttl\|jsonld\|nq`, `-O`

Command	What it does
`tags [type]`	The controlled vocabulary
`tasks`	The curated task pages
`sitemap [kind]`	Every id the hub publishes in its sitemaps
`dump <kind>`	Fetch every record of one kind

Serving

Command	What it does
`serve [--addr]`	Serve the operations over HTTP as NDJSON
`mcp`	Run as an MCP server over stdio
`version`	Print the version and exit

Global flags

These are shared by every operation, so they work the same on every command.

Flag	Meaning
`-o, --output`	Output format: `auto`, `table`, `json`, `jsonl`, `csv`, `tsv`, `url`, `raw`
`--fields`	Comma-separated columns to keep
`--template`	Go text/template applied per record
`--no-header`	Omit the header row in `table` and `csv`
`-n, --limit`	Stop after N records (0 means no limit)
`-j, --jobs`	Concurrent requests
`--token`	Hub token; also read from `$HF_TOKEN`, `$HUGGING_FACE_HUB_TOKEN`, `$HF_HOME/token`
`--deep`	Also read the rendered page and merge the fields only it carries
`--card`	Include the README body on repo records
`--rev`	Branch, tag, or commit sha
`--rate`	Minimum delay between requests
`--retries`	Retry attempts on rate limit or 5xx
`--timeout`	Per-request timeout
`--cache`	Response cache directory
`--no-cache`	Bypass on-disk caches
`--data-dir`	Override the data directory
`--db`	Tee every record into a store (e.g. `out.db`, `postgres://...`)
`-v, --verbose`	Increase verbosity (repeatable)
`-q, --quiet`	Suppress progress output
`--color`	`auto`, `always`, or `never`

See output formats for what -o, --fields, and --template produce, and configuration for environment variables and defaults.