You connect Claude to tools such as HubSpot, Notion and Google Drive through connectors: open the connectors directory in Claude, choose the tool and sign in to grant access. The connecting takes minutes. The decision that matters more is what each connection may do. Start read only, and keep a person approving anything Claude would change.
Key takeaways
- Claude's connectors, many built on the Model Context Protocol, let it read from and act in connected tools.
- According to Anthropic, connecting can give Claude permission to act in a tool, not just read it.
- According to HubSpot's developer documentation, its remote MCP server lets AI tools take actions in a user's HubSpot account.
- A connector usually works with the signed-in user's own permissions, so a broad user account means a broad connection.
- Start read only, grant write access per action only when needed, and have a person approve changes and anything sent.
How do you connect Claude to business tools?
Open Claude's connectors directory, choose the tool, and sign in to that tool to authorise access. Anthropic describes the directory as a way to connect tools in a few clicks. Once connected, Claude can retrieve information from the tool during a chat and, depending on the connector, take actions there.
Anthropic's tutorial on connectors names examples including Google Drive, Gmail, Asana, Notion and Canva. Others are added regularly, so check the directory for what is current. Many are built on the Model Context Protocol, the open standard Anthropic introduced in 2024 and donated to the Linux Foundation's Agentic AI Foundation in December 2025.
How do Claude connectors work?
A connector links Claude to a tool's data and functions. When you sign in, you grant Claude access on your behalf, usually with the same permissions your own account has in that tool. Claude can then search, read and, where the connector allows, create or change things, such as adding a task or a page.
That is the key point. According to Anthropic, connecting gives Claude permission to act in those sources for you, not just read them. If your HubSpot user can edit every record, a connection signed in as you may be able to as well.
HubSpot's own developer documentation describes its remote MCP server as designed for AI tools to take actions within a user's HubSpot account on their behalf. It also powers HubSpot's connectors for Claude, Copilot, Gemini and ChatGPT.
What should Claude be allowed to do in each tool?
As little as each job needs. For most businesses, Claude should start with read access only: reading CRM records, searching the drive, looking up Notion pages. Write access, such as creating records, editing pages or sending messages, should be added one action at a time, with a person approving the result.
| Tool | Sensible start | Add later, with approval |
|---|---|---|
| HubSpot | Read contacts, deals and activity for research and reports | Draft notes or tasks for a person to confirm |
| Notion | Read and search pages | Create draft pages in a specific space |
| Google Drive | Read and search files | Create new draft documents in a set folder |
Two habits help. First, connect with an account whose permissions match the job, rather than an administrator's. Second, prefer actions that create drafts over actions that change or send. A draft can be reviewed; a sent email cannot be recalled.
Security bodies flag the risk of permissions growing over time. The OWASP MCP Top 10, currently in beta, lists privilege escalation via scope creep, noting that loosely defined permissions "often expand over time, granting agents excessive capabilities".
How do you connect HubSpot, Notion and Google Drive safely?
Connect each with an account scoped to what the work needs, check what the connector can do before relying on it, set any available approval settings for write actions, and write down who connected what and why. Review connections every few months and remove ones nobody uses.
A practical sequence for each tool:
- Decide the job. Which tasks need this tool, and do they need to read or write?
- Pick the account. Use one whose permissions match the job.
- Connect. Through Claude's connectors directory, signing in to the tool.
- Check the capabilities. Note which actions the connector can take.
- Require approval for any action that changes data or sends anything.
- Record it. Who connected it, for what, and when to review.
For CRMs in particular, many businesses keep AI read only, because the CRM drives sales follow-up and reporting and a quiet wrong write spreads. The guide to connecting AI to your CRM covers that choice.
What should you connect first?
Connect the source your most frequent AI job depends on, read only. For many teams that is the shared drive, so Claude can read briefs and past work, or the CRM, so it can prepare account research and reports. Add one connection at a time, and only once the job using it is working.
The risk of doing too much at once is real. According to IBM's 2025 Cost of a Data Breach report, 97% of organisations that reported breaches of AI models or applications lacked proper AI access controls. Connections are exactly where those controls live.
Connections give Claude access to data; they do not tell it what is true or how your business does things. That comes from shared company knowledge, which a company brain holds.
Skein sets up connections read only first, with write access granted one tool and one action at a time; the shared knowledge layer page explains the permissions model. That suits a business connecting AI to its CRM, drive or workspace for the first time. A team using Claude only for drafting may not need connectors yet.
Frequently asked questions
How do I connect Claude to HubSpot?
Open Claude's connectors directory, choose HubSpot and sign in with your HubSpot account to grant access. HubSpot's connector runs on its remote MCP server, which can take actions in your account on your behalf, so connect with an account whose permissions match the job and start read only.
Can Claude access Google Drive?
Yes. Anthropic lists Google Drive among its connectors, letting Claude search and read files during a conversation. Connect it through the connectors directory by signing in to your Google account, and consider limiting it to the folders and actions the work actually needs.
Is it safe to connect Claude to our CRM?
It can be, with care. Connectors often act with the signed-in user's permissions, so use an account scoped to the job, start read only, and require a person's approval for any change. Many businesses keep AI read only on the CRM because wrong writes spread quietly through sales processes.
What can Claude do once connected to a tool?
Depending on the connector, Claude can search and read information, link details across tools and take actions such as creating tasks or pages. Anthropic notes that connecting can give Claude permission to act in a tool for you, not just read it, so check each connector's capabilities.
Which tools should I connect to Claude first?
Start with the one source your most frequent AI job relies on, such as the shared drive for briefs and past work or the CRM for account research and reports. Connect it read only, prove the job works, then add the next connection.
Related reading
- Anthropic Academy. "Connect your tools to unlock a smarter, more capable AI companion." Anthropic, accessed 30 October 2026. https://academy.claude.com/tutorials/connect-your-tools-to-unlock-a-smarter-more-capable-ai-companion
- HubSpot Developers. "Remote MCP server." HubSpot, accessed 30 October 2026. https://developers.hubspot.com/docs/build-with-ai/remote-mcp-server
- Anthropic. "Donating the Model Context Protocol and establishing the Agentic AI Foundation." Anthropic, 9 December 2025. https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
- OWASP. "OWASP MCP Top 10." OWASP, accessed 30 October 2026. https://owasp.org/www-project-mcp-top-10/
- IBM. "IBM Report: 13% of Organizations Reported Breaches of AI Models or Applications, 97% of Which Reported Lacking Proper AI Access Controls." IBM Newsroom, 30 July 2025. https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls