This server gives Claude access to BERT-based keyphrase extraction when you need more accuracy than asking an LLM to pull out key terms. It exposes a single tool that reads local files from allowable directories and returns ranked keyphrases without sending the full content to the LLM. Under the hood it runs KeyBERT with spaCy preprocessing and paraphrase-multilingual-MiniLM-L12-v2 embeddings, using MMR to keep results diverse. Reach for this when you're processing documents where bidirectional context matters or you want to save tokens by extracting semantic tags locally before feeding them into your workflow. You can specify stop words and control how many phrases come back.
claude mcp add --transport stdio ivanrublev-keyphrases-mcp -- uvx keyphrases-mcp