Claude HubSpot integration: what actually changes when you connect it

Updated 20 October 2026. HubSpot now applies pipeline stage and conditional property rules when the connector writes, so we corrected that section. We also added the third route, Claude inside a HubSpot workflow, and HubSpot's current read and write lists.

The Claude HubSpot integration most people mean is HubSpot's connector for Claude. It works on every HubSpot plan alongside a paid Claude plan. It reads almost every object in your portal, creates and updates a shorter list, and cannot delete anything. Each person sees only what their HubSpot permissions allow, and every write is logged against them.

One thing has moved since we first wrote this in December 2025. Back then HubSpot's documentation said custom validation rules, including pipeline stage validations, were not applied when the connector wrote to a record. Its documentation, last updated 19 October 2026, now says conditional property rules and pipeline stage validation rules are applied. Association label validations still are not.

The rest of this covers which of the three integration routes you want, what each one can touch, where the limits sit, and what an Australian business is expected to have checked before an AI assistant reads customer records.

Which Claude HubSpot integration do you need?

People say "connect Claude to HubSpot" and mean one of three things. They are different products with different owners.

The HubSpot connector for Claude is the no-code path. HubSpot describes it as available to all HubSpot customers across all tiers who have a paid Anthropic subscription: Pro, Max, Team or Enterprise. Super Admins and users with App Marketplace Access can connect it straight away. Everyone else waits until a Super Admin approves it, and that approval is where you choose which data permissions the connector gets and who in the account may install it. In Claude the path is Settings, then Connectors, then Browse. Search for HubSpot, click Connect, log in to the right portal, tick the permissions you want to allow and click Connect app. To use it in a chat, open the + menu, choose Connectors and switch HubSpot on.

HubSpot's remote MCP server at mcp.hubspot.com is the path for anything you build yourself. It reached general availability on 13 April 2026, it reads and writes live CRM data, it works on every HubSpot plan at no extra cost, and it authorises with OAuth 2.1 with PKCE. We wrote up the rules and limits we put around our own MCP build, and what MCP actually is, in plain terms, if you want the protocol without the marketing. When we build a reporting or data layer for a client we often go one level lower, through a HubSpot private app granted read-only scopes, so the build can see what it needs and change nothing.

Claude inside a HubSpot workflow is the route nobody types into. A workflow reaches a step, sends the record's details to Claude through the Anthropic API, and writes the answer back to a property: a summary of a new lead, a category for an inbound email, a drafted reply that waits for a person. HubSpot's custom code action is how you do this, and it needs Data Hub Professional or Enterprise. The model usage is billed to your own Anthropic API account, and the API key sits in HubSpot's secrets store. It is a build, so somebody has to own the code, the prompt and what happens when a call fails. Before you write one, check whether HubSpot's own Breeze agents already do the job, and read how agentic AI, AI agents and plain automation differ so you build the smallest thing that works.

Which one you want: if a person is going to type at it, the connector. If your own software is going to call HubSpot, the MCP server or the API. If HubSpot should call Claude as one step in a process, a workflow. None of them asks for anybody's HubSpot password. The connector and the MCP server are authorised by logging in to HubSpot and approving what the tool may do, and a workflow runs inside your own portal.

Plan and permission details from HubSpot's knowledge base and developer documentation, checked 20 October 2026.

What can Claude read in HubSpot?

More than most people assume, and worth reading as a list instead of as "our CRM". HubSpot's documented read list is contacts, leads, companies, deals, tickets, custom objects, line items, quotes, quote templates, invoices, orders, carts, payments, payment links, products, subscriptions, segments (lists), campaigns and their attribution data, landing pages, website pages, blog posts, marketing emails, marketing events, pipelines, properties, users, teams, saved reports, buyer intent, engagements including calls, meetings, notes, tasks and emails, the Partner Client Object, and conversations from your Inbox and Help Desk channels.

Check custom objects in your own portal before you rely on them. In August 2026 the connector did not return custom objects for one of our clients, so we used a private app for that job instead. HubSpot's table now lists custom objects as read, create and update, so the gap may have closed, but a test with one real record takes two minutes.

Read that again with your worst-behaved property in mind. Engagements covers call notes and logged emails, which is where people write the things they would never put in a structured field. If your reps record "chasing payment, going to legal" in a note, that is readable. The one exception: if your account has HubSpot's Sensitive Data setting turned on, the connector cannot read any engagement data at all.

What can Claude write to HubSpot, and what are the limits?

Create and update, on a shorter list: contacts, leads, companies, deals, tickets, custom objects, line items, products, quotes and quote templates (both in beta), pipelines, segments, landing pages, website pages, blog posts, marketing emails and engagements. Campaigns, marketing events and properties can be created but not updated. Invoices, orders, payments, subscriptions, users and teams are read only.

Three limits sit on top of that.

1. It cannot delete. HubSpot's wording: "The connector currently supports view, create, and update actions but does not support delete operations." Genuinely reassuring, and it also points the usual paranoia at the wrong risk. The damage an assistant does to a CRM is not deletion. It is a run of records updated slightly wrong, which no restore will ever find, because nothing is missing.

2. Association label validations are skipped. HubSpot's wording now: "When creating or updating records through the connector, conditional property rules and pipeline stage validation rules are applied." And straight after it: "Association label validations aren't applied." That is a real improvement on December 2025, when none of those rules applied. If your reporting depends on association labels, such as a decision maker, a billing contact or a referring partner, check them after any session where Claude has been writing, or keep write access off the objects where those labels matter.

3. Bulk writes cap at ten records. "When creating or updating HubSpot records in bulk, you can only create or update 10 records at a time." The mass tidy-up job people imagine is not this tool. On top of that, "all functionality in Claude is subject to HubSpot's API usage limits and guidelines", so a chatty afternoon counts against the same budget as your integrations.

HubSpot's recommended setting is to make the connector ask before it writes: in the HubSpot connector tool settings in Claude, set Write tools to Needs Approval. Leave it there. Turning it off converts a tool that proposes into a tool that acts, and that is a different risk conversation with a different set of obligations attached. The same setting sits under every skill in HubSpot's Sales plugin for Claude.

Every create and update made through the connector is recorded in your HubSpot Audit Log against both the person and the Claude connector. When a record is wrong, you can see who asked for the change.

Whose permissions does Claude use inside HubSpot?

Straight from the documentation: "The HubSpot connector for Claude automatically respects the user permissions defined in HubSpot. This means users will only see in Claude the CRM data they're allowed to access in HubSpot. For example, individual sales reps will only see pipeline data for deals they have permission to view and update."

That is the best decision in the whole design. It also means your HubSpot permission model has just become your AI data policy, whether or not it was ever built to be one.

There is a second layer on top. When a Super Admin approves the connector, they choose which data permissions it has at all. Widen those later and each person has to disconnect and reconnect before Claude sees the newly approved data, so access only grows when someone decides it should.

Two questions to answer before you connect it:

• Who in your portal has Super Admin, and does the number surprise you?
• Which properties would you be uncomfortable having read aloud in a conversation, and are they actually restricted, or merely unfashionable to look at?

If the honest answer to the second is "everyone can see everything, we just trust people", the problem was already there, and the connector just makes it visible. The fix is HubSpot's own permission settings, including the objects that carry customisable permissions, and the time to use them is before somebody asks Claude which customers are on the biggest discounts.

What does Australian law expect before Claude reads your CRM?

This is the part every walkthrough leaves out. Some of it is the Act and the rest is what the regulator has said it expects, and neither one is a good surprise.

On 21 October 2024 the OAIC published two guides. The one that applies to you is privacy and the use of commercially available AI products. The other, privacy and developing and training generative AI models, is for businesses building and training their own. Three points in the first bear directly on a CRM connector.

The Privacy Act applies to the inputs and the outputs, not only the records you store. The guidance is explicit: "privacy obligations will apply to any personal information input into an AI system, as well as the output data generated by AI (where it contains personal information)". A prompt containing a customer's details is a use of their personal information. So is the summary Claude writes from it, and if that summary gets pasted into a note you have created a new record carrying the same obligations as the original.

Do a privacy impact assessment. The OAIC's stated position is that "organisations considering the use of AI products should take a 'privacy by design' approach, which includes conducting a Privacy Impact Assessment". For a tool that can read every contact record in your portal, that is a short document and an easy one to be embarrassed about not having when somebody asks.

Do not put personal information into publicly available generative AI tools. The OAIC recommends against entering personal information, and particularly sensitive information, into publicly available generative AI tools, because of the privacy risks involved. Which is why the plan you are on is a privacy question before it is a billing one, and why "half the sales team already pastes CRM data into a free chat window" is the risk a properly configured connector reduces instead of introduces. On Anthropic's commercial products, which include Team, Enterprise and the API, Anthropic says it does not use your inputs or outputs to train its models by default. Pro and Max sit under Anthropic's consumer terms, so check the model training setting on any of those accounts before they connect a CRM.

The write side is covered by the ASD's Careful adoption of agentic AI services, published 1 May 2026 jointly with CISA, the NSA and the cyber security centres of Canada, New Zealand and the United Kingdom. Its recommendations map onto this integration almost line for line: adopt agentic AI incrementally and start with low-risk tasks, put approval steps and intervention points on high-risk actions, keep human-in-the-loop review wherever the cost of an error is high, hold strict privilege controls and strong identity management, and keep auditing and reversibility after the task.

The identity point is the one people skip, and it is why we never let an agent borrow a colleague's session. When a record is wrong, the first question is always who changed it. The connector now answers that by logging the person and the connector together. Anything you build yourself has to answer it too, so give each build its own integration user. If the answer is the name of whoever happened to authorise it, you have lost your audit history and put that person in an awkward position.

Two more obligations are better known in advance than discovered. APP 11 requires reasonable steps to protect the personal information you hold, and the Notifiable Data Breaches scheme gives you 30 days to assess a suspected eligible breach. Both are in context in the MCP write-up. There is also a transparency obligation for automated decision making arriving on 10 December 2026, which matters the moment an assistant is making calls about people instead of describing them. Each of these rules, with its date, is set out in our guide to AI governance in Australia.

What is the Claude HubSpot integration good at?

Worth naming, since the limits above take up most of this piece.

• Reading across objects in one question. "Show me every deal over $50k with no activity in three weeks, and what the last email said." That is three reports and a lot of clicking in HubSpot, and one sentence here.
• Writing the note nobody writes. Logging a call, summarising a ticket thread, setting a next step. Engagements are on the write list, and this is the lowest-risk write available.
• Drafting from real context. An email that references what was actually agreed in March, because it read the thread instead of guessing at it.
• Finding the disagreement. Ask which of two records is right and it will tell you they contradict each other, which your dashboards never do.

What it is poor at is anything where being confidently wrong is expensive and nobody checks. That is a property of your process more than of the model. If the job is deterministic and you can write the rule down, a workflow or an integration beats an assistant every time, and the Zapier and HubSpot guide covers that ground. When the job does need judgement on every record and nobody is typing, that is an agent, and our guide to AI agents for Australian businesses sets out what one costs and the controls it needs.

How do you set up the Claude HubSpot integration safely?

1. Approve it with read permissions only, on a sandbox or a low-stakes portal. A Super Admin chooses the data permissions at approval, which takes ten minutes.
2. Ask it the questions your team asks the CRM. "Which deals have gone quiet." "What did we agree with this client in March." "Which accounts have no next step."
3. Write down the three writes you would eventually want, and what a wrong version of each would cost you.
4. Grant write access to the narrowest of the three, with Write tools still set to Needs Approval, and check which association labels those objects rely on, because those are not validated on write.

Step two is the one that surprises people, and it is the reason to do this even if you never grant a write. An assistant reading your CRM out loud is a blunt mirror. Most of what it gets wrong is not the model, it is that two records disagree, an association is missing, or the field three dashboards depend on has been empty since March. If that is what you find, the work is CRM setup and field mapping, and connecting an assistant first only makes the mess faster.

Frequently asked

Do we need a paid Claude plan?

For the HubSpot connector, yes: Pro, Max, Team or Enterprise. The HubSpot side works on every product and tier. The remote MCP server also costs nothing on HubSpot's side, though the model usage behind it is still billed wherever you are calling it from.

Can it delete a record?

No. Create and update only. That is a hard limit of the connector instead of a setting you have chosen, which is useful when a security reviewer asks.

Will it respect our field-level permissions?

It respects the permissions of the user asking, so a person only sees in Claude what they can already see in HubSpot. Which is why the permission model is the thing to fix first if it has drifted.

Where does the data physically go?

That depends entirely on which door you walked through, and it is a longer answer than it sounds: where your data actually goes once Claude is connected.

Is this the same as Breeze?

No. Breeze is HubSpot's own AI, running inside HubSpot on HubSpot credits. This is an outside assistant reading and writing HubSpot through an authorised connection. Exhaust the native tooling before building anything, because a Breeze agent's behaviour is HubSpot's problem and a custom build's behaviour is yours.

Does it apply our validation rules?

Partly. HubSpot says conditional property rules and pipeline stage validation rules are applied when the connector creates or updates a record. Association label validations are not.

Can Claude run inside a HubSpot workflow?

Yes, through a custom code action that calls the Anthropic API and writes the answer back to a property. Custom code needs Data Hub Professional or Enterprise, and the model usage is billed to your own Anthropic API account.

Does Anthropic train Claude on our HubSpot data?

Not by default on Anthropic's commercial products, which include Team, Enterprise and the API. Pro and Max are consumer plans under different terms, so check the training setting on those accounts before they connect.

Where that leaves you

We are a HubSpot Diamond Partner, Certified in the Claude Partner Network, and we build on Claude, which mostly means we have made these decisions on our own portal before making them on anyone else's. How we build AI engineering work like this is the shape of it, our guide to Claude for business in Australia has the plans and what we have built on it, and most of the design work here is deciding what to leave out.

A question to leave you with. If Claude updated one property on every deal in your pipeline tonight, which property would you want a person to have approved first?

Sources

• HubSpot, set up and use the HubSpot connector for Claude
• HubSpot, Claude connector overview
• HubSpot developer docs, custom code actions in workflows
• HubSpot developer docs, integrate with the remote HubSpot MCP server
• Anthropic, is my data used for model training?
• OAIC, privacy and the use of commercially available AI products
• OAIC, privacy and developing and training generative AI models
• ASD ACSC, Careful adoption of agentic AI services
• OAIC, APP 11 security of personal information
• OAIC, Notifiable Data Breaches scheme