davanstrien/hf-find

hf-find is a command line extension for the official Hugging Face hf tool that lets you search for datasets and models on the Hugging Face Hub by meaning instead of exact keywords. Typing a phrase like medieval manuscripts will return datasets that are conceptually related, even if that exact phrase does not appear in their titles, because the tool searches over embeddings and summaries covering roughly 536,000 dataset cards and 636,000 model cards. Once installed with a single hf extensions install command, you get several subcommands: searching datasets, searching models, finding repos similar to a given one, checking what is trending, and viewing the backend's health status. Results come back as a table showing a similarity score, likes, downloads, and a short summary, and you can filter by things like task type, license, programming language, or how recently a repo was modified. Model searches can also be filtered by parameter count, accepting shorthand like 2B or 500M. A notable feature is that hf-find automatically detects when it is being run by an AI coding agent rather than a human, using environment variables set by tools like Claude Code, and switches its output format accordingly. In agent mode it prints clean tab separated data to standard output and sends hints, like suggested follow-up commands, to a separate error stream, so an agent can parse results reliably. There is also a plain JSON output mode for structured integrations. Errors are designed to always include a one-line explanation plus a suggested next step, rather than a raw failure message. For users who want to search or analyze the entire dataset offline rather than one query at a time, the README points to a public Hugging Face dataset containing all the underlying summaries and embeddings, which can be queried directly with a tool called DuckDB.