The short version
- HubSpot’s MCP server went generally available at mcp.hubspot.com in April 2026 - OAuth-based and working with Claude, ChatGPT, Cursor, or any MCP client - alongside a one-click Claude connector and a self-hosted npm package for developers. All three CRM ecosystems we cover now speak the same protocol.
- Three doors in, one rule: the one-click connector for business users, the remote server for teams, the private-app npm package for developers and background agents. Pick by who is connecting, and grant only the scopes the job needs.
- The Smart CRM makes HubSpot unusually good at conversational work: contacts, companies, deals, and tickets share one record spine, so a single conversation can walk from engagement history to pipeline to support without stitching systems together.
- HubSpot adds a twist the other platforms do not have yet: Breeze agents can be MCP clients themselves, calling external tools like Notion, Zapier, and Gong through MCP connectors. The protocol runs in both directions.
- Know the limits before you architect: the remote server does not cover custom objects, MCP calls share your normal API rate limits, and there are no triggers - a deal changing stage cannot wake an agent. Workflows own events; MCP owns judgment.
- Start read-only with scoped access, keep human gates on writes and sends, and audit monthly - the same green-zone discipline that runs through this whole series.
The Trilogy Completes
We have now written this story twice: how AI agents safely reach a Salesforce org, and how they reach a Zoho suite. HubSpot is the third act, and in some ways the most decisive one: in April 2026 HubSpot took its MCP server generally available at mcp.hubspot.com, shipped a one-click Claude connector beside it, and opened its developer platform to custom MCP servers. Every CRM ecosystem we practice on now speaks the same protocol.
That standardization matters more than any single feature. A year ago, connecting an AI assistant to your CRM meant a bespoke integration per platform. Today the question is no longer whether your CRM is agent-ready - Salesforce, Zoho, and HubSpot all are - but how disciplined your setup is. This guide is the HubSpot-specific version of that discipline.
One paragraph of protocol recap for new readers, then we get practical. MCP - the Model Context Protocol - is an open standard that lets an AI client like Claude discover and call tools you have approved. The AI never gets your password or a database connection; it gets a menu of specific, permissioned actions and can order only from that menu. Anthropic introduced it in late 2024, and it has since become the standard rail between AI and business software.
Three Doors Into Your Portal
HubSpot ships three distinct connection paths, and choosing the right one is the first architectural decision.
Door one: the one-click Claude connector. Inside Claude, the connector directory lists HubSpot. Connect, sign in, approve the permissions - done, no configuration. This is the right door for a founder or marketer who wants to ask questions of the portal today. Access follows the signing-in user’s own HubSpot permissions.
Door two: the remote MCP server. Add mcp.hubspot.com to any MCP-capable client - Claude, ChatGPT, Cursor, Windsurf - over OAuth 2.1 with PKCE. Same capability as the connector, but client-agnostic and the standard choice for a team standardizing its tooling.
Door three: the self-hosted npm package. The official @hubspot/mcp-server package runs locally with a private app token you scope yourself under Settings, Integrations, Private Apps. This is the developer door - and the only door for background or scheduled agents, because the OAuth doors require a human completing a browser sign-in.
The rule that spans all three: grant the scopes the job needs and nothing more. A reporting assistant needs crm.objects.deals.read - it does not need write access to contacts, and it absolutely does not need a super-admin token. Scope discipline is the whole security model in three words.
Why the Smart CRM Is a Great MCP Citizen
HubSpot’s architectural bet - one Smart CRM record spine under every hub - pays off unusually well in conversational work. On platforms where marketing, sales, and support live in separate systems, an agent has to hop between tools and reconcile identities. In HubSpot, contacts, companies, deals, tickets, and their entire engagement history already share one graph. One conversation walks the whole customer:
A Monday morning, one prompt - Claude with HubSpot connected
PROMPT: Good morning. Three things:
1. List my open deals with no activity in 14+ days, sorted by
amount.
2. For the top three, summarize the last email thread and draft
a re-engagement note in my voice.
3. Check whether any of those companies have open support
tickets - I do not want to chase a deal while their ticket
sits unanswered.
Show me everything BEFORE logging any tasks or sending anything.
Every part of that prompt rides the shared record spine: deals, engagement history, and tickets resolve to the same companies without any stitching. And the closing clause is the human gate again - reads and drafts run free, writes and sends wait for a person. That discipline is the one constant across all three platforms in this series.
Workflows that earn their keep in week one: the stale-deal sweep above, closed-lost post-mortems ("summarize the last 7 days of closed-lost deals and the stated reasons"), pre-call briefings ("everything we know about this company across marketing touches, deals, and tickets"), and list hygiene checks ("find contacts with no owner and suggest assignments based on territory").
The Twist: Breeze Agents Are MCP Clients Too
Salesforce and Zoho gave us the same two-pattern story: AI inside the platform (Agentforce, Zia Agents) versus the platform inside your AI (MCP). HubSpot has both patterns too - Breeze agents inside, MCP server outside - but it adds a third move neither sibling has shipped yet: Breeze agents can act as MCP clients themselves, reaching out to external tools like Notion, Zapier, and Gong through MCP connectors.
Read that again, because it inverts the usual diagram. In the standard story, your AI assistant is the client and the CRM is the tool. In HubSpot’s version, the CRM’s own agents can also be clients, and the outside world is their toolbox. The protocol runs in both directions through the same standard.
Practical guidance unchanged from the rest of the series: repeatable volume belongs to Breeze (and we activate it properly - data readiness, guardrails, enablement - as a fixed-scope package); judgment, exploration, and cross-record reasoning belong to the conversational side. The connector twist mostly matters if you are building Breeze agents that need your wider stack - suddenly that is a configuration, not an integration project.
The Honest Limits
Three limits shape every serious HubSpot MCP architecture, and none of them are in the launch posts.
No custom objects on the remote server. As of writing, the remote MCP server covers standard CRM objects and engagement history - not custom objects. If your portal leans on custom objects (many industrial and financial portals do), the conversational layer simply cannot see them through the remote path. Workarounds: the self-hosted package plus direct API calls, or a custom MCP server on HubSpot’s developer platform. Decide this before you design, not after.
Rate limits are shared. MCP calls run against the same underlying APIs as everything else - on the private-app path that is on the order of a hundred requests per ten seconds, shared with your integrations. An enthusiastic agent session during a big sync can crowd real automation, exactly as we warned on the Zoho side.
No triggers. The series constant: a deal changing stage, a form submission, a ticket arriving - none of it can wake an MCP agent. HubSpot Workflows own events; MCP owns judgment. The production pattern is a Workflow that notifies a human (or queues work) and a conversational agent that handles the reasoning when engaged.
The Adoption Path: Same Ladder, HubSpot Rungs
Pick one job with real pain
Not "connect AI to HubSpot" - "stop hand-building the Monday pipeline review" or "stop losing closed-lost lessons".
Choose your door deliberately
One-click connector for a person, remote server for a team, private app for developers and background agents.
Scope read-only first
crm.objects read scopes for exactly the objects the job touches. No super-admin tokens, ever.
Run the job conversationally for a week
The same questions you answer by hand today. Note what the agent gets right, wrong, and almost-right.
Add write scopes one at a time
Task creation first - it is low-risk and immediately useful. Notes and records later.
Human gate on anything outward-facing
Emails, deal-stage changes, and anything a customer might see get a person’s approval. Reads and drafts run free.
Wire events through Workflows
Triggers stay in HubSpot Workflows; hand judgment-heavy steps to the conversational layer.
Audit monthly, trim scopes
What was asked, what was done, which scopes were actually used. Unused permissions are risk with no return.
Where This Goes Wrong
- The super-admin token shortcut. It works in the demo and haunts you forever. A scoped private app takes ten extra minutes and removes whole categories of risk.
- Discovering the custom-object gap in week three. If your portal runs on custom objects, the remote server cannot see them. Architect for it up front or lose the rebuild time.
- Treating conversation as automation. The same prompt typed daily is a Workflow or a Breeze agent that has not been built yet.
- Ignoring shared rate limits. Agent sessions and integrations draw from one pool; schedule heavy conversational work away from heavy sync windows.
- No owner. Scopes, connectors, and the audit rhythm need an operator the way a portal needs an admin. Unowned agent setups rot quietly.
What to Do This Week
If your business runs on HubSpot, the first experiment is genuinely free: connect Claude through the one-click connector, ask the stale-deal question, and watch it walk your pipeline. That is thirty minutes including the internal argument about scopes. The self-hosted and remote paths are a Friday afternoon.
And if you want it built properly - scopes designed, the custom-object question answered honestly, Breeze activated where volume belongs and the conversational layer where judgment does - that is our HubSpot practice, delivered as fixed-scope packages from onboarding to Breeze AI Activation to managed services. The series thesis holds on every platform: the protocol is standard now; the discipline is what you are actually buying. Bring the job to a call and we will map it with you.