Hugging Face Hub CLI Skill
Give agents command-line control over Hugging Face Hub models, datasets, Spaces, jobs, and infrastructure.
The evidence shows coverage of downloads, uploads, authentication, cloud jobs, sandboxes, and endpoint management, with some destructive operations exposing --yes, --dry-run, or irreversible-operation warnings. However, the curl-piped shell installer, unclear permission boundaries and sensitive-data flows, token exposure via `hf auth token` stdout, incomplete confirmation guidance, rollback coverage, and extension-install security review reduce trust. The publisher is unverified, but that status alone is not deducted.
Command grouping, parameters, and the replacement of the deprecated command are reasonably clear, and the file identifies generation from `huggingface_hub v1.23.0`. Still, this is only a static command inventory: key-path tests, input-boundary coverage, dependency availability, and diagnosable abnormal-failure guidance are absent. End-to-end reliability therefore remains unproven and is scored conservatively under the static ceiling.
The description covers models, datasets, Spaces, jobs, papers, repositories, and cloud storage, with broad trigger terms. However, triggering on essentially any AI/ML or cloud-storage request is overbroad, while non-fit boundaries, input/output contracts, exclusion conditions, and Chinese-language support are not defined. Core functionality depends on huggingface.co and related overseas infrastructure, whose reachability from mainland China is not addressed, so points are deducted.
The evidence includes Agent Skills formatting, installation instructions, grouped commands, generated-version metadata, README installation paths, an Apache-2.0 license, a security-report channel, and CI validation of generated artifacts. It lacks task-oriented examples, troubleshooting, known limitations, stable-parameter guidance, a changelog, and clearly assigned maintenance responsibility, so it does not merit full marks.
The command inventory directly covers many Hub-management tasks and offers agent/json output formats, so the core CLI value is plausible. However, there are no representative outputs, verified task results, end-to-end workflow examples, or comparison with manual/API alternatives. Complex cloud and destructive operations also require substantial contextual judgment, so the score is capped by static evidence and reduced for incomplete direct usability.
A pinned revision, generated-version note, README generation and CI checks, security policy, and workflow evidence provide some traceability. There is no committed test suite covering the skill's key paths and no third-party execution evidence or independently reproducible results, so verifiability remains limited under static review.
- The installer executes a remote script directly through a curl-to-shell pipeline; pin or verify the script before execution.
- The skill includes token printing, uploads, repository or bucket deletion, job execution, endpoint deployment, and endpoint deletion; require explicit per-action confirmation, least-privilege tokens, and a recovery plan.
- The trigger description is overly broad and may invoke this CLI skill for general AI/ML requests that do not require Hugging Face operations.
- The core service depends on overseas Hugging Face infrastructure; mainland-China reachability, proxy configuration, and offline alternatives are undocumented.
What does this skill do, and when should you use it?
This skill provides operating guidance for the Hugging Face Hub CLI, `hf`, covering authentication, caching, repositories, models, datasets, Spaces, Buckets, papers, jobs, and inference endpoints. It fits users who need agents to perform Hugging Face Hub administration and machine-learning infrastructure tasks from a shell. It also covers discussions, pull requests, collections, webhooks, extensions, and sandboxes. It is one skill in the huggingface/skills monorepo and does not represent the specialized workflows of sibling skills.
Installs and uses the hf CLI; logs in, switches, and manages access tokens; downloads and uploads model, dataset, and Space files; creates, duplicates, moves, deletes, and configures Hub repositories; manages Buckets, caches, collections, discussions, and pull requests; browses models, datasets, Spaces, and papers; runs SQL queries against dataset Parquet URLs; runs, schedules, inspects, and cancels Hugging Face Jobs; creates and manages sandboxes, Spaces, webhooks, and Inference Endpoints; and produces table, JSON, agent, or quiet output.
- A machine-learning developer needs to batch-download or upload model, dataset, or Space files.
- A team needs to manage Hub repositories, branches, tags, visibility, and pull requests.
- A user needs to run or schedule training, inference, or data-processing jobs on Hugging Face infrastructure.
- A researcher needs to inspect dataset Parquet files, execute DuckDB SQL, or verify a local cache.
- An operator needs to deploy, pause, scale, or maintain Hugging Face Inference Endpoints or Spaces.
What are this skill's strengths and limitations?
- Broad coverage of Hub resources, jobs, sandboxes, and inference endpoints.
- Supports JSON and agent-oriented output formats.
- The README says the skill is generated from the locally installed CLI and can be updated with `hf skills update`.
- Explicitly replaces the deprecated `huggingface-cli` command.
- Requires shell, network access, and local filesystem access.
- Many operations require a Hugging Face account or access token; the source does not define exact permission boundaries or costs.
- The source provides no test suite, platform compatibility matrix, or troubleshooting guidance.
- The command surface is large, so first-time users will need `hf --help` for detailed usage.
How do you install this skill?
Install the CLI with: curl -LsSf https://hf.co/cli/install.sh | bash -s. Then place this skill directory in one of Codex's standard .agents/skills locations, or install hf-cli from the huggingface/skills Claude Code plugin marketplace. The README does not provide one universal command that installs this SKILL.md into every client.
How do you use this skill?
In a coding agent with the skill installed, request a concrete Hub operation, for example: “Use the hf-cli skill to log in to Hugging Face, download openai-community/gpt2, and list its files.” Use hf --help for command details; authentication commands are under hf auth, such as hf auth whoami. The source recommends providing credentials through the HF_TOKEN environment variable.
How does this skill compare with similar options?
This skill uses the current hf command and replaces the deprecated huggingface-cli. The source provides no feature comparison with other CLIs or SDKs.