HubSpot MCP and the Claude Connector: What They Can Do With Your CRM Data

Sooner or later, someone asks whether you can connect HubSpot to Claude so it can work directly with your CRM data. Maybe they've seen a demo. Maybe the sales team is already pasting exports into an AI assistant and wants a cleaner way to do it. Either way, the request tends to land with whoever owns HubSpot.

Connecting the two takes only a few minutes. HubSpot now has official connectors for Claude, ChatGPT, Gemini, and Microsoft Copilot. The Claude connector runs on the HubSpot MCP server. The harder part is deciding what the connection should be allowed to see and change. A connected assistant can pull CRM data into a conversation for analysis. Depending on the assistant, it can also create and update records without anyone opening HubSpot. Both are useful. Both also raise access questions worth settling first.

What the HubSpot MCP Server Is

MCP, short for Model Context Protocol, is an open standard for connecting AI tools to your CRM and other business systems. The HubSpot MCP server is HubSpot's own implementation, a hosted connection that lets an AI assistant query live CRM data in plain language and, where you allow it, create and update records. HubSpot moved the server out of beta and made it generally available to all HubSpot accounts in April 2026. That release added write access, engagement history, and marketing content.

The most important design choice is easy to miss. Every action through the server respects the connecting user's existing HubSpot permissions. A sales rep who can only see their own deals in HubSpot sees only those deals through the assistant. The connection doesn't open a back door around the permission model you already have.

HubSpot also has a separate developer MCP server for building apps and CMS assets. If you aren't writing code, it isn't the one you need.

MCP Server or Connector, and Which One You Need

People ask about both but most of the time they're asking about the same thing, just packaged differently.

The MCP server is the underlying connection. A connector is a ready-made way to use HubSpot data from a specific assistant. HubSpot describes its Claude connector as an official MCP connector it built. Once it's installed, Claude uses the HubSpot MCP server to read from and write to your CRM. A separate HubSpot sales plugin adds prepackaged workflow skills on top, without granting any additional access.

There are three ways in. Most teams only need the first.

• Official connector. Built for a specific assistant and approved by a HubSpot admin before the team can use it.
• Custom MCP connector. Created in HubSpot under Development to connect other MCP-compatible tools or custom agents. Usually a developer's or partner's job.
• Developer MCP server. For building HubSpot apps and CMS assets. It isn't a CRM data connection.

Connectors are also different from Breeze, HubSpot's own AI. HubSpot positions Breeze for work inside HubSpot and the connectors for bringing CRM context into the AI tool a team already uses. The line isn't absolute, but it helps when deciding where a task belongs.

What the HubSpot Claude Connector Can Read, Create, and Update

The HubSpot Claude connector is free on HubSpot's side and available on every HubSpot tier. It needs a paid Claude plan (Pro, Max, Team, or Enterprise). Its activity counts against HubSpot's normal API usage limits.

HubSpot's documentation currently describes the access in broad categories.

Read. Nearly anything a user can see in HubSpot. That covers CRM records and custom objects, engagement history (emails, calls, meetings, notes, and tasks), inbox and help desk conversations, marketing performance, commerce objects such as quotes and invoices, and saved reports.

Create and update. CRM records, custom objects, and logged activities. On the marketing side, landing pages, website pages, blog posts, and marketing emails can be created or updated. Pipelines and properties can be created too.

Delete. Nothing. HubSpot states the connector can view, create, and update records but can't delete them. Removing data still happens in HubSpot.

A few limits are worth knowing before anyone relies on it. Bulk changes are capped at 10 records per request. HubSpot's conditional property rules and pipeline stage validations apply to Claude's writes, though association label validations don't. If your account has Sensitive Data turned on, the connector loses access to engagement data entirely. Custom sensitive data properties are never exposed. Custom object support is fairly recent, but it is available.

In our own HubSpot work, that reach was wide enough that the real question quickly shifted from what the connector could do to what we should do through it.

Connecting ChatGPT, Gemini, or Microsoft Copilot to HubSpot

For a HubSpot ChatGPT integration, the official route is the HubSpot connector for ChatGPT, which ChatGPT may list as the HubSpot app. It's the closest match to Claude's. It launched in June 2025 as a read-only deep research connector. In February 2026 it gained the ability to create and update records and log activities. Deep research mode produces longer reports that cite back to HubSpot records. Approval, permission inheritance, the 10-record cap, audit logging, and the no-delete rule work the same way they do for Claude.

Validation is the main documented difference. HubSpot's ChatGPT documentation says custom validation rules, including pipeline stage and association label validations, aren't applied when ChatGPT writes to HubSpot.

Gemini and Copilot are read-only. Google built the HubSpot connector for Gemini, which reads contacts, companies, deals, and their associations. Its availability depends on your Google Workspace setup. The HubSpot agent for Microsoft 365 Copilot is deployed by a Microsoft 365 admin, works in Copilot and Teams, and requires a Copilot license or an eligible Copilot Chat add-on. Either works well for analysis. Changes still happen in HubSpot.

Whichever assistant you choose, the data it retrieves is handled under that provider's terms, so the plan type matters. HubSpot states that Anthropic doesn't train on connector data unless a customer opts in or submits feedback. ChatGPT business plans don't use connector data for training, while consumer plans depend on each user's settings.

In practice, the assistant decision usually follows the company's existing AI standard more than any connector feature. If the team expects to log activity or update records from the assistant, that narrows the choice to Claude or ChatGPT today.

Getting CRM Data Out Without an Export

Until recently, a real answer from the CRM usually meant building a report, asking an admin to build one, or exporting to a spreadsheet. Harder questions went to whoever understood the data model. The connector removes most of those steps. Anyone with access can ask a question in plain language. The assistant pulls the records, sorts them, and helps interpret the result.

That self-serve access shows up in everyday work. A sales leader pulls a pipeline summary by stage and amount, sorted by close date, before a forecast meeting. An account manager reviews a customer's full conversation history and open issues before a call. A marketer asks which campaigns or email subject lines performed best last month.

At Ascendex, this is where we've used the connector most, partly because HubSpot's interface already handles most of the data entry we've encountered well. Plenty of teams lean on write access more, which is also a valid use case.

Answers still deserve a check, since the connector sits alongside your reporting, not in place of it. Before a number goes in front of leadership, compare it with the records and with a HubSpot report you already trust. HubSpot's own guidance is to ask the assistant for citations back to the records it used. Clean data improves results, though it doesn't have to be perfect. Messy data also tends to surface faster once more people can ask questions of it.

Writing to HubSpot Through the Connector

Connector writes save real steps when the work is already happening in the assistant. The useful question for any task is whether it belongs in the assistant or in HubSpot's own tools.

These tend to be good fits for writing through the connector:

• Logging a call summary or notes right after a conversation
• Creating follow-up tasks from a meeting or an email thread
• Updating a deal stage or a few properties during a pipeline review
• Creating a contact or deal from details already in the conversation
• Drafting a marketing email or landing page in the assistant, then finishing it in HubSpot

Other work is usually better handled in HubSpot directly.

• Bulk updates. The 10-record cap applies per request, so a string of requests can still change a lot of records. Imports and workflows are built for volume and are easier to check.
• Anything that should happen every time. That's a workflow, not a prompt.
• Final review and publishing. The connector can publish blog posts, landing pages, and marketing emails. Layout, review, and approval live in HubSpot, so decide deliberately where publishing happens.
• Changes that depend on validation rules. Neither connector applies every rule HubSpot's own interface enforces. The gaps differ by assistant.

Every create and update through the Claude and ChatGPT connectors is recorded in HubSpot's audit log, attributed to both the user and the connector. That trail makes a mistake easier to trace after the fact. Catching it before it happens is where permissions come in.

Two Layers of Permissions, and Who Controls Each

Connector permissions sit in two places. An admin needs to set both.

On the HubSpot side, Super Admins and users with App Marketplace Access can connect without prior approval. Everyone else needs a Super Admin to approve the connector first. During approval, the Super Admin chooses which optional data permissions the connector gets and who in the account can install it. If those permissions change later, each user reconnects to pick them up. Underneath all of it, the connector respects each user's existing HubSpot permissions, down to record-level restrictions.

On the assistant side, each connector tool can run automatically or stop and ask first. HubSpot recommends setting write tools to "Needs Approval" in Claude. ChatGPT offers the same kind of setting. With "Always allow," edits can happen without a confirmation step. That setting doesn't grant any new rights in HubSpot. What it removes is the moment where a person confirms the change.

That confirmation matters more than it might seem. OpenAI's security guidance describes prompt injection, where instructions hidden in content an assistant reads can steer what it does next. CRM emails and notes are exactly that kind of content. No HubSpot-specific incident has been documented but keeping confirmations on for writes is still a simple guard.

Together, the two layers are what make the connector reasonable to turn on. HubSpot decides what the connection can ever do. The assistant setting decides what happens automatically and what waits for a person. Each deserves a deliberate choice, not a default.

One thing to watch out for applies to almost every account. Anyone with Super Admin or App Marketplace Access can connect a connector, including outside agencies and consultants. Review who holds those permissions, and check connection activity in the audit log. On Claude Enterprise, admins can also stop users from connecting HubSpot to personal Claude accounts outside the company workspace.

What to Decide Before You Turn It On

Most of these decisions take one short conversation. They're much harder to unwind after dozens of people have connected.

1. Who gets access first. A small pilot group without admin rights is easier to learn from than the whole company.
2. Which data. Decide which optional data permissions to approve, and whether any objects or sensitive fields should stay out.
3. Which writes to allow. Some teams start read-only. Others start with one narrow write use case, such as logging notes or tasks. Either way, try writes on test records before live ones.
4. Approval settings. Make "Needs Approval" for write tools a team standard, not something each person decides alone.
5. Which assistant and plan. Confirm the plan's data terms, and remember that only Claude and ChatGPT can write to HubSpot today.
6. Who owns it. Someone should review connection activity and connector writes in the audit log, track new capabilities as HubSpot adds them, and know how to reverse a bad change.

These sit on top of the broader readiness questions that apply to any CRM, which we laid out in what has to be true before you connect AI to your CRM.

Start With One Team and a Few Real Tasks

The connector itself is the easy part. The value comes from picking a small group and a handful of real tasks, then settling the permission and approval standards that fit your team. Expand from there as people learn what works.

Ascendex helps companies put AI to work across their CRM and the rest of the business with the right access, approvals, and ownership in place from the start. If you're working out how HubSpot and an AI assistant should fit together, talk with our CRM team about setting it up so the answers hold up and the changes stay under control. Contact us to start the conversation about your next technology initiative.