Happy slice of mcp pizzaMCP.pizza

mcp-hubspot

MCP.Pizza Chef: baryhuang

Sixteen tools, well over twice what the write-up lists: reading, creating and updating both contacts and companies, pulling recent conversations and support tickets, and reading or changing the custom fields your records use. You connect it with a token you create inside HubSpot yourself. Two catches worth knowing. The permissions the setup guide tells you to grant do not cover the ticket or custom-field tools, so those fail until you add more. And the built-in search only looks through records already fetched, never HubSpot itself.

Communication
Data

Use This MCP server To

Look up a contact before a call without leaving chat Turn a pasted profile into a new contact record See what has been happening with a company lately Update a contact's job title after a promotion Read recent support conversations for one customer Check which support tickets are still sitting open

README

HubSpot MCP Server

Docker Hub License: MIT

Overview

A Model Context Protocol (MCP) server that enables AI assistants to interact with HubSpot CRM data. This server bridges AI models with your HubSpot account, providing direct access to contacts, companies, and engagement data. Built-in vector storage and caching mechanisms help overcome HubSpot API limitations while improving response times.

Our implementation prioritizes the most frequently used, high-value HubSpot operations with robust error handling and API stability. Each component is optimized for AI-friendly interactions, ensuring reliable performance even during complex, multi-step CRM workflows.

Why MCP-HubSpot?

  • Direct CRM Access: Connect Claude and other AI assistants to your HubSpot data without intermediary steps
  • Context Retention: Vector storage with FAISS enables semantic search across previous interactions
  • Zero Configuration: Simple Docker deployment with minimal setup

Example Prompts

Create HubSpot contacts and companies from this LinkedIn profile:
[Paste LinkedIn profile text]

What's happening lately with my pipeline?

Available Tools

The server offers tools for HubSpot management and data retrieval:

Tool Purpose
hubspot_create_contact Create contacts with duplicate prevention
hubspot_create_company Create companies with duplicate prevention
hubspot_get_company_activity Retrieve activity for specific companies
hubspot_get_active_companies Retrieve most recently active companies
hubspot_get_active_contacts Retrieve most recently active contacts
hubspot_get_recent_conversations Retrieve recent conversation threads with messages
hubspot_search_data Semantic search across previously retrieved HubSpot data

mcp-hubspot FAQ

What do I need to connect it?
An access token generated from a private app inside your own HubSpot account. It is not your password, and free HubSpot accounts can create one.
Can it change or wipe my records?
It can create and update contacts, companies and the custom fields your records use, and it does all of that without asking you to confirm. It has no delete tool, so nothing can be erased outright.
The setup lists three permissions. Is that enough?
Not for everything it can do. Tickets and custom fields need permissions that are missing from that list, so those tools return a refusal until you widen the token yourself.
Can I use this to add a contact from a profile I paste in?
Yes, and it is the author's own headline example. It checks for an existing match first so you do not end up with two copies of the same person.
What does the built-in search actually search?
Only the records this tool has already pulled down in earlier sessions. It never asks HubSpot anything new, so a name it has not seen before simply will not come up.
Where does my customer data end up?
Contact details and conversation text are copied into a searchable store on your own machine. If you skip the folder-sharing line the setup marks as optional, that store is thrown away every time you close it.
How hard is setup?
It runs as a container, so you need Docker installed and comfortable, plus a first run that downloads a language model before anything answers.
Is it still maintained?
The last change was November 2025, so it has been quiet for several months, though the code itself is more complete than its own documentation admits.