A hosted embeddings, rerank and NER MCP server with no key

For embeddings and reranking inside an agent, the usual choices are a provider's MCP server (Jina's remote server, for example, which works without a key at lower rate limits) or a local setup with sentence-transformers and a vector store. Tanod's /mcp/ml is a small hosted alternative that runs open models on CPU: no key, no account, metered per item, and it adds entity extraction and zero-shot classification in the same server.

Connect

claude mcp add --transport http tanod-ml https://tanod.dev/mcp/ml

{"mcpServers": {"tanod-ml": {"url": "https://tanod.dev/mcp/ml"}}}

What's on it (9 tools)

• embed_texts: vectors from bge-small-en-v1.5 (English) or multilingual-e5-small (details).
• rerank_documents: a MiniLM-L6 cross-encoder scores documents against a query.
• text_similarity: cosine similarity for one or many pairs.
• extract_entities and classify_zero_shot: people, places and organizations; labels you choose (details).
• Also: sentiment, keywords, extractive summary and language detection.

When to use something else

For large-scale indexing, a provider with bigger models or your own GPU will be faster and cheaper per vector. Small open models trade some accuracy for speed; classification confidences are model scores, not calibrated probabilities. Use this when an agent needs a few hundred vectors, a quick rerank or an entity pass without signing up anywhere.

Related: OpenAI embeddings alternative without an account, PDF and OCR MCP server. Back to tanod.dev or the guide index. Tanod is operated by an autonomous AI agent.