Claude, ChatGPT and your HubSpot data: the connectors explained | Discover Digital

HubSpot's connectors put your CRM inside the AI tools your team already uses: Claude, ChatGPT, Gemini and Microsoft Copilot. They now create and update records as well as read them, so the settings you choose before anyone connects matter more than they did a year ago.

One server, four connectors

HubSpot offers connectors for ChatGPT, Claude, Gemini and Copilot (HubSpot AI tools), and all four run on the same remote HubSpot MCP server. MCP, the Model Context Protocol, is a common way for AI tools to connect to other software. In HubSpot's words, its "own AI connectors for Claude, Microsoft Copilot, Gemini, and ChatGPT are powered by this server" (HubSpot developer documentation).

The connectors are not new. HubSpot launched a deep research connector for ChatGPT in June 2025 and a connector for Claude in July 2025, and announced an updated ChatGPT connector with OpenAI on 16 September 2026. In the Fall 2026 Spotlight, the connectors and the MCP server are both listed as generally available, along with access to conversations (chat, email, SMS, WhatsApp and Messenger) in Claude and through the MCP server (HubSpot Spotlight). This article concentrates on Claude and ChatGPT, where HubSpot's documentation is fullest.

What they can read, create and update

In both Claude and ChatGPT, the connector can read, create and update contacts, leads, companies, deals, tickets, custom objects, line items and products, and log calls, meetings, notes, tasks and emails. Both also work with campaigns and landing pages. Invoices, orders, carts, subscriptions, segments, users and teams are read only (Claude connector, ChatGPT connector). Three limits apply to both:

• Nothing is deleted. "The connector currently supports view, create, and update actions but does not support delete operations."
• Ten records at a time. Bulk creates and updates are limited to 10 records.
• Workflows stay with you. Neither connector lists workflows among the things it can build. HubSpot says that when asked for one, "Claude will provide the drafted content for you to implement in HubSpot."

The differences are in the detail:

• Website pages and quotes. Claude can create and update website pages, and quotes in beta. In ChatGPT both are read only.
• Validation. Claude's writes follow conditional property rules and pipeline stage validation, though not association label validation. In ChatGPT, "Custom validation rules (including pipeline stage validations and association label validations) aren't applied", so a stage rule your process relies on will not stop a change made there.
• Content. Claude sees published pages and posts in full, but only metadata for drafts. It can also read conversations from Inbox and Help Desk.
• Plans. On the HubSpot side, both connectors work on all products and plans. Claude needs a paid Anthropic plan (Pro, Max, Team or Enterprise) and works on the web, desktop and mobile. ChatGPT needs an OpenAI account, and the connector may appear there as "HubSpot App".

Permissions, approval and the audit trail

Both connectors respect each user's HubSpot permissions, so nobody sees or changes more through an AI tool than they could in HubSpot itself. Super Admins can connect straight away, and for Claude so can users with App Marketplace Access. Everyone else waits until a Super Admin approves the connector, chooses its data permissions and decides who may install it.

Every create and update is recorded in HubSpot's Audit Log against both the user and the connector. The connectors cannot see Sensitive Data properties, and if Sensitive Data is turned on in your account they see no engagement data at all.

One rule matters for agencies and groups: you can connect only one HubSpot account at a time to a Claude or ChatGPT account. A consultant who works across several client portals has to choose which one the AI tool is looking at.

Where your data goes

For HubSpot accounts hosted in the EU, connector requests are routed through the EU data centre. Once data reaches the AI tool, that provider's terms apply. HubSpot's documentation says Anthropic does not train on data shared through the connector except where a customer gives feedback or opts in. For ChatGPT, connector data is not used for training on Enterprise, Business or Edu plans; Free, Go, Pro and Plus users can opt out, and otherwise OpenAI "may use connector data to improve connector performance". HubSpot may also collect anonymised summaries of MCP tool calls.

That makes the plan your team uses part of the data protection decision, not just a budget line. Claude Enterprise admins can also restrict connections to accounts on verified domains.

For teams building their own agents: the MCP server

If you are building your own agent, you connect to the same server, using OAuth with PKCE, after creating an MCP connector under Development in HubSpot (HubSpot's developer guide). Its documented abilities include creating, editing and publishing landing pages, website pages and blog posts, drafting marketing emails, and managing campaigns, pipelines, custom properties and static segments, with quotes in beta. It reads conversations from live chat, team email, WhatsApp, SMS, Messenger and custom channels, and respects inbox restrictions. Its limits are worth knowing before you design anything:

• Writes need explicit confirmation from the user.
• Search uses HubSpot's CRM search, with no vector search, and returns up to 200 results a page.
• The query tool has no joins, subqueries or association traversal, so a question that spans contacts, companies and deals takes several steps.
• Sensitive Data blocks conversation data.
• The website page, blog post and preview tools work only with Claude, ChatGPT and Gemini.

HubSpot's Spotlight card for the server mentions workflows, but the developer documentation does not list workflow writes, so we would not plan around that yet. The server is also separate from HubSpot's local Developer MCP server, for building apps and CMS assets, and from Agent CLI, a public beta for bulk record work from the terminal.

A safe way to roll the connectors out to a team

1. Start small. Approve the connector for a few people, set its data permissions and limit who may install it. Check their HubSpot permissions first, because the connector inherits them.
2. Read before you write. For the first few weeks, use it for questions, summaries and reports, not edits. Pipeline reviews and meeting preparation are good first tasks.
3. Set write tools to "Needs Approval". HubSpot recommends this for Claude; with "Always allow", edits can happen without confirmation.
4. Check the Audit Log. Each week, look at what was created and updated through the connector.
5. Choose plans with data in mind. Prefer ChatGPT's business plans, or make sure anyone on a personal plan has opted out of training.
6. Test ChatGPT in a UK account. HubSpot's ChatGPT article no longer mentions regional limits, though earlier versions reportedly did, and OpenAI's help pages are reported to say some apps may not be available in the EEA, the UK or Switzerland. The article's FAQ also contradicts its own table on write access, so confirm what works in your account before you brief the team.
7. Write a one-page guide. What to ask, what never to paste in, the 10-record limit, and who to call when something looks wrong.

We set up the connectors for HubSpot clients and agree with them what each tool can see and change. If you are planning a roll-out, see our AI services for HubSpot.