Hugging Face CLI

Downloads individual files or entire repository snapshots from Hugging Face Hub with built-in resumable downloads, automatic local caching, and offline-mode support. Uses a content-addressable cache architecture where files are stored by their SHA256 hash, enabling deduplication across multiple model versions and automatic cache invalidation when remote files change. Implements HTTP range requests for resume capability and metadata-driven cache validation without re-downloading unchanged files.

Uploads files and entire folders to Hugging Face Hub repositories using either Git-based commits (for version control) or direct HTTP uploads (for simplicity). Automatically handles Git Large File Storage (LFS) for files exceeding size thresholds and supports Xet deduplication for efficient storage of similar files. The commit API abstracts away Git complexity while maintaining full version history and branching support, allowing developers to upload without managing local Git repositories.

Converts models between formats (PyTorch to ONNX, TensorFlow to SavedModel, etc.) and applies quantization techniques (int8, int4, float16) for model optimization. The conversion system integrates with Hub repositories, enabling one-command conversion and re-upload of optimized models. Supports framework-specific conversion pipelines and automatic format detection.

Implements Model Context Protocol (MCP) server for integrating Hugging Face Hub operations into Claude and other MCP-compatible applications. Exposes Hub functionality (search, download, upload, inference) as MCP tools that can be called by LLMs, enabling natural language interaction with Hub repositories. The MCP server handles authentication, request routing, and response formatting transparently.

Manages community features on Hub repositories including discussions, pull requests, and comments. Enables programmatic creation and management of discussions for model feedback, pull requests for collaborative improvements, and comment threads for community engagement. Integrates with repository operations for seamless collaboration workflows.

Creates, deletes, and configures Hugging Face Hub repositories programmatically with fine-grained control over visibility (public/private), access permissions, and metadata. Supports branch and tag management, repository settings updates, and community features like discussions and pull requests. The HfApi class provides a unified interface for all repository operations, handling authentication and error states transparently.

Lists and searches models, datasets, and spaces on Hugging Face Hub with filtering by task, library, language, and other metadata attributes. Returns structured metadata including model cards, download counts, and community metrics. The search API uses Hub's backend indexing to enable fast filtering across thousands of repositories without downloading metadata locally.

Executes inference on 35+ ML tasks (text generation, image classification, object detection, etc.) across multiple providers including Hugging Face Inference API, Replicate, Together AI, Fal AI, and SambaNova. The InferenceClient abstracts provider differences behind a unified task-based API, handling authentication, request formatting, and response parsing. Supports both synchronous and asynchronous execution with streaming for long-running tasks.

Creates, edits, and publishes structured model cards (documentation) for models, datasets, and spaces with YAML frontmatter metadata and markdown content. The ModelCard class provides a programmatic interface for managing model documentation including model architecture, training data, intended use, limitations, and bias/fairness considerations. Cards are stored as README.md files in repositories and automatically parsed/validated against Hub schemas.

Provides ModelHubMixin pattern for integrating any ML framework (PyTorch, TensorFlow, JAX, scikit-learn, etc.) with standardized save/load methods that automatically upload to Hugging Face Hub. The mixin pattern allows frameworks to define custom serialization logic while inheriting Hub integration, enabling one-line model uploads without manual Git/HTTP handling. Supports framework-specific implementations (PyTorchModelHubMixin, TFModelHubMixin) with optimized serialization for each framework.

Provides HfFileSystem, a PyArrow-compatible filesystem interface that treats Hugging Face Hub repositories as mounted filesystems. Enables standard filesystem operations (ls, cp, rm, open) on Hub files without downloading entire repositories, supporting streaming reads and lazy evaluation. Integrates with PyArrow and pandas for direct data loading from Hub without intermediate downloads.

Provides utilities for inspecting, validating, and cleaning up the local Hub cache directory. The cache system uses content-addressable storage (SHA256-based) with metadata tracking, enabling detection of corrupted files, orphaned cache entries, and unused models. Supports cache scanning to estimate disk usage and selective deletion of specific models or entire cache cleanup.

Manages Hugging Face API authentication via tokens stored in local configuration files or environment variables. Supports interactive login flows, token validation, and automatic credential injection into API requests. The authentication system integrates with all Hub operations transparently, handling token refresh and expiration.