Hugging Face skill: what hf-cli does and how to install it

Source: huggingface/skills · Maintainer: community: Hugging Face · License: Apache-2.0 · Checked: 24 October 2026

What is the Hugging Face skill?

The most-installed Hugging Face skill is hf-cli: a generated reference of every hf command, so Claude Code, Codex, Cursor or another agent can search and download models, manage datasets, buckets and Spaces, read papers and run jobs on the Hugging Face Hub without guessing flags. Hugging Face's own README calls it "the recommended first Skill to install".

It lives in huggingface/skills, Hugging Face's official skills repo, alongside about two dozen workflow skills for training, evaluation, Gradio and SageMaker deployment. The README describes the split: the plugin marketplaces "expose the hf-cli skill as the bootstrap path for core Hub operations", and the rest install on demand.

The first sentence of the skill's description reads: "Hugging Face Hub CLI (hf) for downloading, uploading, and managing models, datasets, spaces, buckets, repos, papers, jobs, and more on the Hugging Face Hub." The rest of the description lists its triggers and ends: "Replaces the deprecated huggingface-cli."

As of 24 October 2026, its skills.sh page listed about 3,300 installs (snapshot, 24 October 2026), the most of any skill in the repo. That count only covers installs through the skills CLI, not the CLI installer or plugin routes below.

What does it do?

The skill is model-invoked and casts a wide net. Its description tells the agent to use it whenever you mention "hf", "huggingface", "Hugging Face" or "huggingface-cli", or anything in the Hugging Face ecosystem and AI and ML in general, "even if the user doesn't explicitly ask for a CLI command."

The SKILL.md is a command list, not a workflow. It opens with the CLI install line, notes that auth commands now live under hf auth, and states it was generated with huggingface_hub v2.3.0. Then it lists each command with its flags:

• Top level: hf download, hf upload, hf cp, hf sync, hf env, hf update, hf version.
• Accounts and files: hf auth (login, switch, whoami), hf cache (list, prune, rm), hf buckets (create, sync, remove).
• The Hub: hf models, hf datasets (including hf datasets sql to query parquet files with DuckDB), hf spaces, hf repos, hf collections, hf discussions, hf papers (search, read as markdown), hf webhooks.
• Compute: hf jobs (run and schedule jobs), hf endpoints (Inference Endpoints), hf sandbox (experimental sandboxes on Jobs).
• Tooling: hf extensions and hf skills (add, list, preview, update).

It closes with common options (--format json, -q, --revision, --token, --type), a section on hf-mount for mounting repos and buckets as local folders, and tips: use hf <command> --help, authenticate with HF_TOKEN, and update with hf update.

The folder contains only SKILL.md. There are no scripts.

How do you install it?

Hugging Face documents several routes in its CLI for AI agents guide and the repo README. Pick one.

1. The standalone CLI installer (adds the skill for you). The hf CLI guide says the installer "also installs the hf-cli skill globally, for Claude Code and any agent reading ~/.agents/skills":

curl -LsSf https://hf.co/cli/install.sh | bash

Pass --exclude-skill to skip it (bash -s -- --exclude-skill). On Windows the guide gives a PowerShell installer instead.

2. hf skills add (if you installed the CLI with pip or Homebrew). From the agents guide:

hf skills add --global   # all projects
hf skills add            # current project only

3. Claude Code plugin. From the agents guide:

/plugin marketplace add huggingface/skills
/plugin install hf-cli@huggingface-skills

The repo README writes the second line as hf-cli@huggingface/skills; the marketplace's own manifest names it huggingface-skills, which matches the docs. The plugin route installs the skill but not the CLI itself, so install hf too.

4. Skills CLI. The command shown on skills.sh:

npx skills add https://github.com/huggingface/skills --skill hf-cli

That copies the folder to .claude/skills/hf-cli/ in your project, or ~/.claude/skills/hf-cli/ with -g, the standard Claude Code skill locations. Note that this copy is the repo's pre-generated file, not one built from your CLI version.

Other agents. The README lists the Cursor Marketplace, copying folders into .agents/skills for Codex, and a Gemini CLI extension (gemini extensions install https://github.com/huggingface/skills.git --consent). Codex and Cursor read the same SKILL.md format; see Codex skills and Cursor skills.

More Hugging Face skills. The marketplaces expose only hf-cli. For the rest, the README says to run hf skills add <skill-name>; hf skills list shows the catalog. On skills.sh, the next most installed were huggingface-datasets, huggingface-gradio and huggingface-llm-trainer, each between about 2,200 and 2,400 installs (snapshot, 24 October 2026).

How do you use it?

Log in once with hf auth login (or set HF_TOKEN), then ask in plain language.

Example prompts (suggestions, not quotes from the repo):

• Find the five most downloaded text-embedding models under 500M parameters and show their licenses.
• Download only the safetensors and config files from Qwen/Qwen3-0.6B into ./models.
• Upload ./checkpoints to a new private model repo called my-org/support-classifier.
• Run a SQL query on the train split of this dataset and count rows per label.
• Search papers on the Hub about speculative decoding and summarize the top three.

Requirements: the hf CLI on your PATH and, for anything beyond public reads, a Hugging Face account and token. Commands like hf jobs run, hf endpoints deploy and hf sandbox create start compute on Hugging Face infrastructure under your account, so check costs on huggingface.co before letting an agent run them.

Tips and limits

From the files and pages linked, not from testing:

• Security audits on skills.sh. As of 24 October 2026, hf-cli showed Pass from Gen Agent Trust Hub and Socket and Warn from Snyk, all audited 24 October 2026. Snyk's note was a medium third-party content exposure risk, because hf discussions list lets the agent read discussions and pull requests written by outsiders.
• Keep it in sync with your CLI. The agents guide says the generated skill "goes stale when you upgrade". hf update refreshes a globally installed skill and hf skills update -g does it alone. hf reminds you once a day; HF_HUB_DISABLE_UPDATE_CHECK=1 silences that.
• It triggers broadly. The description asks to be used for anything AI or ML related, so it may load in sessions where you only mention a model. That costs context but runs nothing by itself.
• Tokens. The skill says to prefer the HF_TOKEN environment variable over --token, and the CLI guide recommends the environment variable to keep tokens out of your command history.
• Old names on skills.sh. Listings such as hugging-face-cli and hugging-face-model-trainer also appear on skills.sh with far fewer installs; the current README uses the hf-cli and huggingface-* names.
• License. Apache-2.0, so you can copy and modify it with the license notice kept.

Save it in your Kipt library

In Kipt, open Skills and paste the hf-cli folder link into the import box. For a repo with many skills, Kipt shows a list to pick from, so you can also take a workflow skill such as huggingface-gradio. You can also drop the folder or a .zip. Your skills live in your own private GitHub repo and install into Claude Code or Codex with one command:

npx skills add https://github.com/<you>/<repo>.git --all -g

Because hf-cli is generated from a CLI version, re-import it after a big hf upgrade. Kipt is not affiliated with Hugging Face.

Similar skills

• Find skills skill: search skills.sh from inside your agent for more ML skills.
• Claude API skill: Anthropic's reference skill for building on the Claude API.
• MCP builder skill: if you'd rather wrap the Hub in an MCP server than a CLI skill.