HubSpot Connector for Claude: The Operator's Guide

The HubSpot connector for Claude is HubSpot's official integration for Anthropic's Claude, built on HubSpot's MCP server. It lets Claude read your CRM, marketing, and engagement data, then create and update records with your approval. It works on every HubSpot plan at no extra cost; you need a paid Claude plan (Pro, Max, Team, or Enterprise).

What is the HubSpot connector for Claude?

The HubSpot connector for Claude was the first CRM connector Anthropic shipped, launched with HubSpot in July 2025. It sits in Claude's connector directory, and once connected, Claude can answer questions from your live portal data: summarize deals by stage, chart campaign performance, pull the full conversation history with a contact before a call.

It started read-only. Since late 2025 it also writes: creating contacts and deals, updating records, and logging notes, tasks, and meetings straight from the chat window. And since mid 2026 it runs analytical queries against your CRM with a SQL-based retrieval method, so aggregations and cross-object questions come back computed rather than guessed.

Under the hood it runs on HubSpot's MCP server, the same plumbing developer tools can connect to directly. The connector is the packaged, no-code version of that access, and it is part of the same AI push as HubSpot's Breeze agents (our guide to the Prospecting Agent covers that side). And the pattern is bigger than HubSpot's own tools: FirstTouch gives Claude LinkedIn outreach over the same MCP standard, with every action logged back to HubSpot.

What can Claude read and write in HubSpot?

Access is per object, and it is broader than most people expect. Here is HubSpot's permission model, condensed:

Nothing can be deleted, on any object. Custom objects, unstructured data, and custom Sensitive Data properties are not available. Campaign attribution data is read-only. The Partner Client Object is read-only and only surfaces when your prompt names it explicitly.

How do you set up the HubSpot connector for Claude?

Setup takes about five minutes. The order matters because the first connection sets permissions for everyone who follows:

1. Admin connects first. A Super Admin or a user with App Marketplace permissions opens Claude, goes to Settings, then Connectors, browses to HubSpot, and adds it to the team.
2. Choose the portal and the permissions. Click Connect, log in to the HubSpot account you want, and select which permissions the connector gets. Then click Connect app. Trimming permissions later means uninstalling and reinstalling, for the admin and every user, so decide the scope here.
3. Grant your team access. In HubSpot: Settings, then Integrations, then Connected Apps, then HubSpot connector for Claude, then Give users access.
4. Each user connects their own Claude account. They get only the permissions the admin selected, further limited by their own HubSpot user permissions.
5. Turn it on in chat. Open Search and tools in the Claude chat window and toggle HubSpot on. While you are in the tool settings, set Write tools to Needs Approval.

How do permissions and security work?

The core model is simple: Claude inherits the HubSpot permissions of whoever connected. A rep sees in Claude exactly the deals they can see in HubSpot, nothing more. Every create and update lands in the portal's audit log, attributed to both the user and the connector, and connection events are logged too.

On the data side: requests route through HubSpot's data centers in your portal's region (EU portals stay on EU routing), and after that your Anthropic terms govern handling. Anthropic does not train on data shared through the connector, except in cases like a user choosing to give feedback or opting in. If your portal has Sensitive Data turned on, the connector gets no engagement data at all, by design.

Three one-way doors to know before you connect. Giving users access cannot be undone, even if you uninstall. Once users connect, admins cannot revoke the selected permissions; narrowing scope means uninstall and reinstall for everyone. And uninstalling the connector does not disconnect your users; each has to uninstall individually. Our take: the user-level permission model is well built, and the admin controls are the thinnest part of the whole integration. Decide your scopes before the first connect, not after.

What are the limits to know before you connect?

The honest list, from HubSpot's own documentation and from running this daily:

• Custom validation rules are not applied. Pipeline stage validations and association label validations do not run on connector writes. A rep can move a deal in Claude in a way the HubSpot UI would block. This is the single biggest thing to understand before granting write access.
• The bulk cap is a speed bump, not a guardrail. Bulk creates and updates are limited to 10 records at a time, but admins on HubSpot's community report that Claude will happily loop through batches of 10 if asked. Approval mode is your real control.
• No custom objects. Custom objects, unstructured data, and associated engagements are not yet available, and the leads object is absent from the supported list.
• One portal at a time. A Claude account connects to a single HubSpot portal. Agencies and partners juggling portals disconnect and reconnect to switch.
• Published content only. For website pages, landing pages, and blog posts, only published data is available; drafts expose metadata only.
• No deletes, anywhere. Honestly, a feature. Claude cannot remove records, engagements, or values on your behalf.

How Hubjoy runs HubSpot through Claude

We operate HubSpot portals through Claude and AI agents every day: CRM operations, pipeline hygiene, reporting. This list comes from daily use, not a demo.

1. Scope the connection like a new hire. The connector inherits the connecting user's permissions, so the user IS the permission model. Connect with an account whose access matches what Claude should see, not a Super Admin's view of everything.
2. Write tools on Needs Approval, permanently. Every create and update gets a human eye before it lands. Because validation rules do not run on connector writes, treat every write as if the UI guardrails were off. They are.
3. Reads before writes. Run it as a question-answering and reporting layer for a couple of weeks first. You learn which properties it reads reliably and where your data lies to it.
4. First writes go to low-stakes fields. Notes and tasks before deal stages. Test records before your best accounts. Widen access only after the output earns it.
5. Check the audit log weekly. Connector writes are attributed to both the user and the connector, which makes a weekly filter cheap insurance against surprises.
6. Data hygiene still decides quality. Claude answering from stale stages and duplicate contacts is confidently wrong. Clean the portal first; the connector amplifies whatever state your CRM is in.

The gotcha we plan around: the one-way doors. Access granted to users cannot be ungranted, and narrowing permissions means everyone uninstalls and reinstalls. Five minutes of scope planning before the first connect saves a very awkward cleanup later.

Frequently asked questions

Using ChatGPT instead of Claude, or both? HubSpot ships an equivalent connector for OpenAI's assistant - see our guide to the HubSpot Connector for ChatGPT for its plans, security model, and where the two connectors differ on write access.

Need to connect a tool other than Claude or ChatGPT, like Cursor or a custom agent? Both packaged connectors run on HubSpot's own MCP server - see our guide to the HubSpot MCP server for the generic, developer-facing setup.

Connecting AI to your portal?

Claude is only as good as the CRM it reads. We will tell you whether your portal is ready to be operated by AI, and what to fix first if it is not. Thirty minutes, no pitch.