HubSpot MCP server: what it does and what it can't see

The HubSpot MCP server is HubSpot's official implementation of the Model Context Protocol: it lets AI assistants like Claude or ChatGPT read and act on your CRM data (contacts, companies, deals, tickets, engagements) through natural language, under the connected user's permissions. Ask "summarize my open deals by stage" or "create a follow-up task for every contact who opened yesterday's email" and the assistant executes against your live portal. This guide covers what the server does, the three ways to connect it, its documented limits, and the blind spot that determines how useful it actually is.

‍

What is the HubSpot MCP server?

MCP (Model Context Protocol) is the open standard that lets AI applications talk to external systems through typed tools rather than custom integrations. HubSpot ships two distinct servers, documented on its developer MCP page. The remote HubSpot MCP server (mcp.hubspot.com) gives any MCP-compatible AI client secure read and write access to CRM data via OAuth: contacts, companies, deals, tickets, engagements, and a growing set of marketing objects. The Developer MCP server is a different animal: it runs locally through the HubSpot CLI and serves developers building apps or CMS content on the platform. For querying your CRM from Claude, the remote server (or the connector built on it) is the one you want.

What you can do with it

Once connected, the assistant turns the CRM into a conversational surface. The recurring use cases:

• Pipeline analysis: "summarize active deals by amount and stage, sorted by close date", "which deals have had no activity in 14 days".
• Record operations: create and update contacts and deals, log notes, tasks and activities without leaving the chat.
• Engagement context: pull the email, call and meeting history behind a record before a conversation.
• Reporting shortcuts: ad hoc questions that would otherwise mean building a report or exporting to a spreadsheet.

The documented limits matter for planning, per HubSpot's own connector documentation: bulk operations are capped at 10 records per request, custom validation rules are not applied on writes, and sensitive data properties are excluded from access. MCP is built for conversational, ad hoc work; bulk jobs and webhooks remain API territory.

How to connect Claude to HubSpot

Three paths, by profile:

• 1. Official connector (most teams). Add HubSpot from the connector directory in Claude, authenticate with OAuth, done. No code. Requires a paid Anthropic plan (Pro, Max, Team or Enterprise) and admin approval on the HubSpot side.
• 2. Remote MCP server (custom clients). Create a user-level app with the scopes you need on the HubSpot Developer Platform, then point any MCP-compatible client at mcp.hubspot.com with the OAuth credentials.
• 3. CLI and community packages (developers). The HubSpot CLI exposes an MCP setup command, and open-source packages accept a private app access token for self-hosted control.

One piece of advice whatever the path: the assistant inherits every permission of the connected user. Scope read-only first, verify the outputs against your portal for a week, then grant write access object by object.

The limit nobody mentions: what is actually in the fields

The MCP server is a pipe, not a source. Ask Claude "which deals in negotiation have an unresolved pricing objection" and it will faithfully query your deal records, then answer from whatever the objection field contains, which in most portals is nothing. Budget, decision process, competitors in the deal, methodology criteria: these live in conversations, and they only exist in HubSpot if someone (or something) put them there. Connecting an AI to a CRM full of empty strategic fields produces confident answers about incomplete data. Before expecting intelligence from the pipe, fix the source: our guide to CRM data quality covers exactly that problem.

Pair it with conversation data

Praiz closes that gap from both sides. First, it feeds HubSpot: configurable AI agents run on every sales and customer conversation (recorded and transcribed in 100+ languages) and write the extracted output into individual HubSpot properties through the native integration: pain, budget, MEDDIC or BANT criteria, objections, competitor mentions, next steps. Every HubSpot MCP query instantly gets smarter, because the fields it reads are actually filled. Second, Praiz has its own MCP server, which exposes the conversation layer itself: from Claude or ChatGPT you can query transcripts, scores and signals across your whole call base, then cross them with HubSpot data in the same chat. "Which negotiation-stage deals had a pricing objection in the last call, and what did the rep answer" becomes a real, answerable question.

Praiz customer teams measure (internal data) a CRM completion rate multiplied by 5, +90% reliability on strategic CRM fields, and 100% of deals with MEDDIC completed automatically versus 32% before. One all-inclusive plan at €30 per user per month (annual) on the pricing page, with hands-on onboarding and agent configuration included: the property mapping that makes MCP queries useful is set up with you, not left to your admins.

Frequently asked questions