Give Claude Code the Entire HuggingFace Ecosystem: Three Commands to Unlock Model Search, Training Jobs, and Hub Pushes

Three Claude Code commands wire in the entire HuggingFace Hub — search models, launch GPU fine-tuning jobs, push checkpoints, and manage datasets from a single plugin marketplace install.

Claude Code is great at writing code. HuggingFace Skills makes it great at running ML workloads — searching the Hub, submitting training jobs to cloud GPUs, evaluating models, and pushing finished checkpoints, all without leaving your Claude Code session.

Installing the HuggingFace Skills plugin

HuggingFace publishes a plugin marketplace at huggingface/skills. Add it to Claude Code and install the HuggingFace CLI skill:

# 1. Register the marketplace (one-time)
/plugin marketplace add huggingface/skills

# 2. Install the hf-cli skill (teaches Claude every hf command)
/plugin install hf-cli@huggingface/skills

# 3. (Optional) Install the fine-tuning skill
/plugin install hf-llm-trainer@huggingface/skills

The hf-cli skill is generated from your locally installed HuggingFace CLI, so it's always current. Once installed, Claude Code understands every hf subcommand — search, datasets, models, Spaces, jobs, and buckets.

What Claude Code can now do

You: Search for Llama-3.2 fine-tunes that support 128K context and are < 8B params

Claude: [searches Hub] Found 14 models matching those criteria:
        - meta-llama/Llama-3.2-8B-Instruct (8B, 128K) ✓
        - unsloth/Llama-3.2-8B-Instruct-bnb-4bit (4-bit quant) ✓
        ...

You: Fine-tune the 8B Instruct model on datasets/my-org/support-tickets
     with LoRA on a single H100, push to my-org/support-llama

Claude: [launches job on HuggingFace Jobs, monitors progress, pushes on completion]

With the hf-llm-trainer skill, Claude understands the full training decision space: which GPU tier to pick for a given model size, LoRA vs full fine-tuning, SFT vs DPO vs GRPO, how to configure Hub auth, and how to structure multi-stage pipelines.

Available skills

Skill	What it teaches Claude
hf-cli	All hf CLI commands: Hub search, datasets, models, Spaces, jobs
hf-llm-trainer	Fine-tuning from 0.5B to 70B params; SFT, DPO, RLVR; GGUF conversion
hf-datasets	Dataset loading, filtering, format conversion, push to Hub

Install on demand with /plugin install <skill>@huggingface/skills. Each skill is a SKILL.md file with YAML frontmatter — you can read and customize them.

Why this matters for your workflow

The HuggingFace ecosystem is where open-model AI engineering happens: model discovery, dataset curation, training, evaluation, and serving. Skills make all of that available inside the same Claude Code session where you're writing application code — no context switch to a notebook, no separate CLI window, no copy-paste of model IDs.

HuggingFace published a worked example of using this to have Claude fine-tune an open-source LLM end-to-end — from dataset prep to Hub push — without the developer writing a single training script: We Got Claude to Fine-Tune an Open Source LLM.

Sources: HuggingFace Skills docs — huggingface.co/docs/hub/en/agents-skills · huggingface/skills — GitHub · We Got Claude to Fine-Tune an Open Source LLM — HuggingFace Blog · Claude Code Integration — DeepWiki