What people mean by a HubSpot MCP server that respects user permissions
They mean two different worries at once. The first is that an AI assistant will see records the person asking could never open in HubSpot. The second is that it will delete or merge something, and nobody will be able to say who asked for it.
The usual trigger is a rep who has already connected something. A sales manager wants deal data in Claude. An operations lead wants contact lists in ChatGPT. Somebody sets it up, it works, and then the security review asks whose HubSpot seat the AI was using. If the answer is "a shared private app token," the review stops there.
MCP is the protocol an AI client speaks to reach tools. Both answers below use it. They differ in what sits between the client and HubSpot.
HubSpot's own remote MCP server: what HubSpot documents
HubSpot runs a remote MCP server and hosts it itself, and documents it on its own developer site (HubSpot MCP server, checked November 2026). HubSpot says the server graduated from beta and is now generally available. It says any MCP-compatible AI tool can read from and write to a HubSpot CRM through it, and that all actions through its MCP server respect the user's existing HubSpot permissions. HubSpot describes the server as a HubSpot-hosted connection authenticated with OAuth 2.1 and PKCE, and says connecting starts with creating an MCP auth app from inside the HubSpot account (HubSpot changelog, checked November 2026).
That is a real per-person permission model, and it is HubSpot's own. The permissions it leans on are the ones a HubSpot admin already set per user (HubSpot user permissions guide, checked November 2026). If every person on the team has a HubSpot seat, and the only app anybody needs is HubSpot, this is the shorter path. There is no second vendor, no second bill, and no second place to look when something breaks.
HubSpot also documents one restriction. If a HubSpot account has sensitive data protection turned on, Activity objects are blocked from HubSpot's MCP server specifically. Activity objects means calls, meetings, notes, tasks and emails. That restriction does not apply to HubSpot's standard CRM APIs (HubSpot changelog, checked November 2026).
What Elaichi's HubSpot connector adds
Elaichi is a governed MCP control plane. Its HubSpot connector is one Elaichi authors and runs itself, which is a different thing from HubSpot's own server. It connects with OAuth and carries 284 tools, covering contacts, companies, deals, properties and workflows.
The per-person part works the same way, with the same mechanic. Each rep connects their own HubSpot account. Every call through a connection runs on the account that connection was made with. A connection a rep made themselves acts as that rep's HubSpot seat, with whatever that seat can see.
Three things sit on top of that, and they are the reason to add a layer at all.
One address for every client. Claude, ChatGPT, Cursor and Microsoft Copilot all sign in to the same Elaichi endpoint, https://api.elaichi.ai/mcp. There is no per-person URL, no per-team URL, and nothing with a token in it. Which organization and which person a client reaches comes from the OAuth grant behind its token.
Tool-level rules an admin writes once. A restriction in Elaichi names connectors and tools, and targets either a role or one member. That is how you keep a destructive tool away from a whole team without touching anybody's HubSpot seat.
One audit log across apps. Every connected-tool call that reaches execution writes a row. The row records the person, the tool, the connector, the connection it reached, what it touched and which MCP client made the call. HubSpot, Slack, Jira and the rest land in the same log, filterable in one place.
Which HubSpot tools to block, and for whom
Start with the ten that are destructive or hard to undo. In Elaichi's HubSpot connector these are:
delete_a_hubspot_contact_by_idhubspot_contacts_mergehubspot_contacts_batch_deletedelete_a_hubspot_company_by_idhubspot_companies_mergehubspot_companies_batch_deletedelete_a_hubspot_deal_by_iddelete_a_hubspot_workflow_by_iddelete_a_hubspot_property_by_iddelete_a_hubspot_user_by_id
These are not all the same. The connector's own descriptions say the contact, company, deal and property deletes archive the record, recoverable for a window of 90 days, and the batch company delete behaves the same way. The workflow delete is the permanent one. delete_a_hubspot_user_by_id removes a HubSpot user, which is an admin action rather than a CRM edit and does not belong on a sales role at all.
A merge belongs on the list beside the deletes. Merging two contacts is not reversible from the AI's side. A model that cannot tell two similar records apart is exactly the thing that proposes a merge.
Write one block rule on the Member role naming the HubSpot connector and those tools. If you would rather not maintain a list, the restriction editor's "Block deleting" preset ticks every tool on the connector that deletes, which covers the deletes without naming each one; the merges are not deletes, so add them by hand. A block rule matches by name or by the operation the tool performs, so renaming a tool does not slip it past the block. Reads and ordinary updates keep working, because a layer holding only block rules passes everything it does not name.
If you want tighter than that, an allow rule is the other shape. Be careful with it. An allow rule turns on an allowlist for its whole layer just by existing. A role with one allow rule naming HubSpot read tools is then blocked from every other connector until you add allow rules for those too. Read the difference between an MCP tool allowlist and a blocklist before you pick.
One honest caveat on timing. A restriction change takes effect within about two minutes, not on the next call.
How a rep's call is checked, step by step
A call from Claude through Elaichi to HubSpot passes several gates. Every one of them runs on the person who asked, rather than on whoever set the toolbox up.
The grant's scopes come first. Running any connected tool needs the mcp:tools scope. The person ticked that as "Use your connected apps" on Elaichi's consent screen when they connected their client. The person's role is checked next: without the tool:execute permission their tool list is empty.
Then restrictions. Elaichi resolves the caller's role rules and any rules aimed at them personally. A tool blocked for that person is left out of what the AI client can even find, and refused if it calls the name directly. The refusal tells the caller whether the block comes from their role or their own account, and that an admin can lift it through an access request.
Last, the connection. The call runs on the HubSpot account behind the connection, and HubSpot applies its own permissions to that account. Two layers, both narrowing: Elaichi decides which tools the person may call, HubSpot decides what that seat may see.
Every call that reaches execution writes an audit row. A call refused at the scope check writes none, because that refusal happens before any executor runs.
When HubSpot's own server is the right answer
Stay with HubSpot's own remote MCP server if all of this is true:
- HubSpot is the only app the team needs in their AI client.
- Everybody who needs it already has a HubSpot seat, and HubSpot's own permissions are set correctly on those seats.
- Nobody is asking for a cross-app audit trail, or for a record of AI activity that covers more than one vendor.
- You are comfortable that per-tool limits, if you need them, will come from HubSpot's side rather than yours.
That is a legitimate setup, and adding a control plane to it buys nothing. The case for a layer in front starts when the second app arrives. One address instead of one per vendor, one place to write a rule, one log to filter at the next access review. Vendor MCP servers per SaaS versus one control plane works through that trade-off in more depth.
Rolling it out for a sales team
The order that holds up:
First, pick who connects what. Each rep connects their own HubSpot account, so each person's calls use their own seat. A connection shared with a team runs on its owner's account instead. That is what you want for a shared marketing account, and not what you want for reps. The same split is worked through for Salesforce, per rep.
Second, write the block rule on the Member role naming the delete and merge tools. Check it from a rep's own client afterwards, not from an admin account.
Third, have each person connect their AI client once. Claude, ChatGPT and Cursor all take the same address. Elaichi's sign-in needs dynamic client registration, PKCE with S256 and a browser sign-in. Any client that supports those should connect; test the one your team uses before you tell everybody it works.
Fourth, open the audit log after the first week and see what people actually called. That is usually where the next rule comes from.
One thing a control plane does not solve. Elaichi's MCP endpoint has no prompt-injection defense of its own. A tools/call carries only a tool name and arguments, never the person's prompt. That defense belongs to the client. Restrictions are what bound the damage: a tool that is blocked cannot run, however the model was persuaded to ask for it.
For the full sales rollout, including toolboxes and what to share, see HubSpot MCP access for a sales team, per rep. If marketing is the team asking, should marketing get HubSpot inside ChatGPT covers the campaign side. The connector itself is on the connectors catalog.