Two things are slow with HubSpot: getting records in from unstructured text, and finding a conversation you half remember. Creation handles duplicates itself, and everything retrieved is indexed for semantic search — so the second time you ask, the answer is already local.
A HubSpot CRM client focused on the operations that come up daily: creating contacts and companies with duplicate prevention, reading recent activity, and searching what has been retrieved. Retrieved data is indexed with FAISS, so the search is semantic and survives between sessions.
- Contacts and companies created, with duplicate prevention on both
- The most recently active companies and the most recently active contacts
- Activity for a specific company
- Recent conversation threads with their messages, indexed thread by thread
- Semantic search across everything previously retrieved, backed by a FAISS vector store
- Storage that persists between sessions, in a directory you configure
A HubSpot access token with the CRM scopes the tools use. It deploys as a Docker image, with images built for several architectures; SentenceTransformer produces the embeddings, and caches them automatically.
One command plus a key — npx -y @smithery/cli@latest install mcp-hubspot --client claude, then supply credentials