Connecting Claude to HubSpot without losing control of your data

August 26, 2026 marks a major step forward in collaboration between CRMs and LLMs. Salesforce and Anthropic announce Claudeforce, Salesforce's CRM accessible directly in Claude, with 37 ready-to-use sales skills and a beta opening in September. A turning point in the SaaS world, which accepts no longer bringing users into its native interface. Marc Benioff sums up the shift in one phrase: “the UI is the AI”. The interface is no longer an application you open, it is a conversation. A CRM without a screen, queried from the outside: the market already has a name for it, the headless CRM.
HubSpot is following the same path, more quietly. Since April 2026, its MCP server has made it possible to read from and write to the CRM from Claude. AI is therefore already at the door of your portal. What remains is to decide what you let it do, and where your customer data goes.
At Markentive, the question has been settled for a while: connecting Claude to HubSpot is a given. We do it every day, on our own portal and for our clients. Never without safeguards, however. And this is precisely what separates a gadget from a lever.

Concrete uses

These uses are not theoretical. We practice them every week on our own portal; here are a few examples.
Pipeline review in one question. “Which deals have a passed close date, no activity for three weeks, or an unusual amount?” The answer comes back as a usable list, with no report to build and no export to rework. For us, it has become a Friday ritual: Claude screens the pipeline and only surfaces the anomalies that need a decision. The rest of the time, we no longer think about it.
Diagnosing your configuration. Ask Claude to describe the actual use of your properties, stages and workflows, then compare it with the assumed configuration. The gap is the real deliverable, and it is always wider than the team imagines. We have packaged this audit so it can be rerun from one portal to another: properties never filled in, dead workflows, ghost pipelines. It is the starting point of any CRM integration.
Preparing a meeting. Contacts and roles, email history, latest call notes, open deals on the account. The brief fits on one page and cross-references sources nobody has the time to cross-reference by hand. You walk into the meeting informed, without having opened five tabs.
Reporting and data entry. Notes and meeting summaries created on the fly after a conversation, monthly portfolio summary, follow-ups detected. The most tedious part of the CRM disappears from teams' schedules, just like the follow-ups automated in marketing automation workflows.
Agents that work without you. At Markentive, one bot links HubSpot and Pennylane for invoicing, another monitors the portal and alerts us when a communication goes out to a client. Nobody in front of the screen, and the work is done by morning.
The list is far from closed. Each team discovers its own uses after a few weeks, and that is a good sign: it is the moment the tool becomes a habit.

How the connection works

Three things to know, and none requires a technical background.
The connection uses a standard called MCP, for Model Context Protocol. Think of it as a cable: it is what links Claude to HubSpot, and the whole governance question depends on how this cable is laid.
The shortest path is the official HubSpot connector, available in Claude. Two minutes to activate, no technical project, and everything just described works. It is the natural entry point, the one we recommend starting with: read-only, with a limited group of users, and accounts whose HubSpot permissions are clean.
There is a second path: an MCP server built specifically for your organization. We come back to it below, because this is where control is decided.
In passing, do not confuse this with Breeze. Breeze is HubSpot's built-in AI, a copilot and agents that live inside the portal. Here we are talking about the opposite movement: connecting an external AI, Claude, to your HubSpot data.

What about governance?

Everything above makes you want to open the floodgates. This is exactly the moment to slow down for two minutes, because two questions decide what comes next. Who can do what, and where the data goes.

Who can do what

The connector inherits the permissions of the connected user, with no intermediate setting available to an administrator. A sales rep who can edit deals in HubSpot can therefore edit them through Claude. To start in read-only mode with a limited team, this framework is enough. To roll it out widely, it quickly falls short: it is impossible to give the AI a scope different from the user's existing role in the instance. For an organization still choosing its tool, the granularity of the permission model deserves a place in the CRM decision grid.
A clarification, because the concern comes up often: no, the connector does not allow records to be deleted. The tools exposed cover reading, creating and updating. The real risk is bulk writing without review, and the lack of a log to understand afterwards why 400 deals changed stage one Tuesday evening.

Where the data goes

When a user queries the CRM through the connector, the relevant data is sent to the model: names, emails, amounts, conversation histories. Three facts to frame the debate calmly.
What is covered: on its enterprise offers, Anthropic does not train its models on customer data by default, signs a data processing agreement (DPA) within the meaning of Article 28 of the GDPR and governs transfers through the European Commission's standard contractual clauses. The fantasy of “the model learning from your data” does not survive a reading of the contracts. On the other hand, there is no direct European hosting: the infrastructure of Claude Enterprise and the API is in the United States, and running in the EU requires going through AWS Bedrock or Google Vertex AI in a European region.
What remains structural: Anthropic is a US company, and therefore subject to the CLOUD Act, which allows US authorities to demand data regardless of where it is located. For most companies, this is a theoretical risk long accepted with Google Workspace or Microsoft 365. For some industries, it is a decision criterion in its own right.
What is coming: most of the AI Act has applied since August 2026, and the CNIL, the French data protection authority, has selected recruitment, including automated screening of applications, among its three priority audit themes for 2026. It presents these audits as a preview of its supervisory role under the AI Act. The question “which customer data do we send to which AI provider, and who approved it” will eventually be asked. Better to answer it before it comes from a regulator, a key account customer or your DPO. There is already a precedent on a related topic: the CNIL recommendation on email open tracking forced teams to review their HubSpot campaigns.

Keep the value, regain control

This is exactly the need a custom MCP meets: a server built for your organization, which you control, without losing any of the uses described above.
The tools exposed are chosen one by one: removing write access to deals is not just a denied permission, the action no longer even exists among the model's options. Permissions are set finely, independently of the user's role. Scheduled agents become possible, without anyone in front of a consent screen. And every call is logged: request, tool called, records affected.
Above all, this is the level at which confidentiality is handled. A custom MCP can pseudonymize data before it reaches the model. This is the choice we made for our own use: Markentive's internal MCP replaces personal data with neutral identifiers, and Claude analyzes a pipeline, detects duplicates or prepares a report without ever receiving a client's name. The value of the analysis is intact, and the sensitive data does not leave.
And contrary to what people imagine, deployment is nothing like an IT project. The server is installed as an app in HubSpot and distributed to teams through the company's Claude organization, like any connector. The investment is in the design (which tools, which permissions, which anonymization rules), not in the plumbing.
For those who want to take the logic all the way, there remain open-source models and hosting in your own infrastructure. Relevant for a minority of highly constrained contexts, at a cost of control far beyond what precedes. There are steps before getting there, and most companies will never need to take them all.

Where to start

With our clients, the rollout always follows the same thread.
Audit first. A portal whose data is approximate does not become queryable: it becomes queryable in a confident but wrong way. The instance audit (data, properties, workflows, segments) tells you what needs fixing before connecting anything. The underlying topic remains structuring, covered in our article on enriching CRM data without polluting your database.
Open read access to a limited group. A few weeks are enough to see the questions teams really ask, and to spot where the answers are wrong. Each wrong answer points to a structuring defect: the list becomes a backlog.
Open write access in stages. Activities before deals, notes before pipeline stages. And always someone who reviews.
Train and log. Adoption cannot be decreed: teams need to know what they can ask the tool, and the administrator needs to be able to say who changed what, and why. This is what makes the setup last over time.

AI audit and scoping for your HubSpot portal

Want to know what AI can bring to your portal, and within what framework? We audit your instance and scope the rollout with you: priority uses, permissions, data governance.
Let's discuss your project