REPO PACKAGE
huggingface skills
OFFICIAL26skills
5groups
3ready to use
LABSCO SUMMARY
Hugging Face's 26 skills listed here are a working toolkit for anyone building against the Hugging Face Hub, not internal conventions for contributing to Hugging Face's own codebase.
Three are the general on-ramp to the Hub: hf-cli wraps the hf CLI for downloading and uploading models, datasets, and Spaces; hf-mcp wraps the Hugging Face MCP server for search and repo lookups; and hf-mem estimates how much memory a given Safetensors or GGUF checkpoint needs before you try to load it. A larger group is task-specific — building or hosting a demo (huggingface-gradio, huggingface-spaces, huggingface-zerogpu, huggingface-lora-space-builder), training or evaluating a model (huggingface-llm-trainer, huggingface-vision-trainer, train-sentence-transformers, trl-training, huggingface-community-evals), running one locally (huggingface-local-models, transformers-js), or working with data, models, and papers on the Hub (huggingface-datasets, huggingface-best, huggingface-papers, huggingface-paper-publisher). The last two, huggingface-tool-builder and huggingface-trackio, cover writing custom Hub-API scripts and tracking a training run.
This is for anyone building against the Hub — training a model, hosting a demo, browsing a dataset, or picking one off a leaderboard — regardless of whether they've ever touched Hugging Face's own repositories. Ten of the twenty-six need an account before they do anything real: the LLM and vision trainers, paper-publisher, trackio and zerogpu need a Hugging Face token, and the five hf-cloud SageMaker skills need an AWS account. The rest work against public Hub data or your own machine.
READ THE FULL ANALYSIS
The install path assumes a plugin-capable agent. Hugging Face ships this as a plugin marketplace — /plugin marketplace add huggingface/skills, then /plugin install <skill>@huggingface/skills — rather than a single skill folder you drop in place. Codex, Gemini CLI, and Cursor each get their own install path in the readme; if your tool supports none of them, the repository's fallback is a generated AGENTS.md bundle instead.
What the rest need. Seven need only a local tool already on your machine (hf-cli's own binary, llama.cpp for local-models, and similar), three are ready to use with no setup at all, and five more — hf-mem, huggingface-lora-space-builder, huggingface-spaces, train-sentence-transformers and transformers-js — have you install a package first (uv, huggingface_hub, sentence-transformers or @huggingface/transformers), the Spaces and training ones with an hf auth login as well. One, hf-mcp, is a live MCP dependency rather than a fallback path: it only works with the Hugging Face MCP server configured as a tool source.
ALSO IN THIS PACKAGE
PLUGINS
1 itemPLUGIN
hf-cli
Execute Hugging Face Hub operations using the hf CLI. Install additional Hugging Face skills, discover agent resources, download models/datasets, upload files, manage repos, and run cloud compute jobs.
MCP SERVERS
1 itemMCP SERVER
huggingface-skills
Search and use the Hugging Face Hub from the chat
http · https://huggingface.co/mcp?loginWHAT'S INSIDE
26 showing · 26 totalReady to use
Nothing else to set up — install it and go.
26 of 26
SETUP
Showing all 26 skills
hf-mcp
Connects an assistant directly to Hugging Face, so it can look things up there and run work on its machines while you stay in the conversation.
Choosing a model and running it yourself4 skills
hf-mem
AI models are big files, and one too big for your computer's memory will not run — this works out how much memory a given model needs, without downloading it first.
huggingface-best
Ask which AI model is best at a job and get back a short ranked list, scored on public tests and narrowed down to the ones your own computer can actually run.
huggingface-local-models
Runs an AI model on your own computer instead of somebody else's, in a version squeezed small enough to fit the memory you actually have.
transformers-js
Runs AI models in plain JavaScript, in the browser or on a Node server — no separate Python service to stand up, and in the browser the user's data never leaves their machine.
Demo apps on Hugging Face Spaces4 skills
huggingface-gradio
Puts a web page in front of a piece of Python code, so anyone can try it in a browser — a box to type in, a button to press, a chat window.
huggingface-lora-space-builder
Takes a LoRA — the small add-on file that teaches an image or video generator one particular style or subject — and puts it online as a demo anyone can try in a browser.
huggingface-spaces
Hugging Face will host a small web app for you, graphics card and all; this is how to build one, ship it, and keep it running.
huggingface-zerogpu
On Hugging Face's shared GPUs, the graphics card only exists for the seconds your function is actually running — and almost every strange bug on that hardware traces back to it.
Deploying to Amazon SageMaker6 skills
hf-cloud-aws-context-discovery
Before any work in your Amazon Web Services (AWS) account, this reads which account, region and sign-in your computer is already set up with, changing nothing, and warns early if your login cannot create the permissions a deployment will need.
hf-cloud-python-env-setup
Gives the scripts that deploy to Amazon SageMaker a separate Python of their own, at a version machine-learning libraries support, because most deployment failures that look like AWS problems are really Python problems.
hf-cloud-sagemaker-deployment-planner
The starting point for putting an AI model online on Amazon SageMaker: it asks what the model is and how often it will be called, recommends a setup and machine, and waits for your yes before anything that costs money is created.
hf-cloud-sagemaker-iam-preflight
Every SageMaker deployment needs a permission role in your AWS account. This finds one you already have before trying to make a new one, and if your login cannot create roles, tells you exactly what to ask your admin for.
hf-cloud-sagemaker-production-defaults
The step that actually puts the model online on Amazon SageMaker, with automatic scaling, alarms and cleanup labels switched on from the start, then sends one real request before calling it done.
hf-cloud-serving-image-selection
Picks the software package (container) that will run your model on Amazon SageMaker and finds its current address in AWS's catalog, Hugging Face's own containers first, because the wrong or outdated one is the most common reason a deployment will not start.
Training and testing models6 skills
huggingface-community-evals
Scores an AI model on the standard public exams — reasoning, maths, general knowledge — using the graphics card in your own machine.
huggingface-llm-trainer
Take a ready-made language model and retrain it on your own examples so it picks up your task or your style — the training runs on Hugging Face's machines, and the finished model is saved to your account.
huggingface-trackio
Watches a model while it trains — the numbers arriving on a live chart, and a message when something starts going wrong.
huggingface-vision-trainer
Teaches a model to recognise things in pictures — where an object is, what it is, or its exact outline — on rented cloud machines rather than your own.
train-sentence-transformers
Trains a model that judges how close in meaning two pieces of text are — the thing a search box needs to return the right result for a badly worded query.
trl-training
Takes a general-purpose language model and teaches it your task or your taste, through a handful of command-line commands rather than training code.
Working on the Hugging Face site5 skills
hf-cli
The command line for Hugging Face, the site where the world's open AI models and datasets are kept — one command for the things you would otherwise do by clicking around its website.
huggingface-datasets
A way to look inside a Hugging Face dataset without downloading it, and to put one of your own up there: everything the dataset page on the website shows you, reachable from a script.
huggingface-paper-publisher
Gets your research paper listed on Hugging Face, with your name claimed as its author and the models and datasets it produced linked to it.
huggingface-papers
Hand it a link to an AI research paper and it brings back the paper itself to read, along with who wrote it and which code, models and data go with it.
huggingface-tool-builder
A way to stop doing the same Hugging Face lookup by hand every time: you say what you want to pull out, and a small script is written for you to keep and rerun.
No skill in this package matches that.
HOW TO GET IT
SINGLE SKILLInstall a skill by name
npx skills add huggingface/skills --skill <name> --full-depthPick the skill name from the Skills tab — each entry there installs independently.
ANTHROPIC'S MARKETPLACEInstall the whole set
/plugin install huggingface-skills@claude-plugins-officialTyped inside Claude Code, not in a terminal. Claude Code adds this marketplace by itself the first time you open it; if it is missing, add it first with /plugin marketplace add anthropics/claude-plugins-official. The command opens the plugin's details; confirm there to install. This copy is Hugging Face's set as it stood on 17 September 2026.
HUGGING FACE'S MARKETPLACEAdd Hugging Face's marketplace, then install hf-cli
/plugin marketplace add huggingface/skills/plugin install hf-cli@huggingface-skillsTyped inside Claude Code; the marketplace is called huggingface-skills, which is the part after the @. No MCP server comes with this route. Hugging Face's other skills are added later, one at a time, with hf skills add <skill-name> in a terminal, once Hugging Face's hf command-line tool is installed.