Deletes and bulk changes wait for a person.

Guardrails out of the box

The Data Hub classifies every call as read, write or destructive. Deleting, archiving, removing and unsubscribing count as destructive. It detects them from the tool, the arguments and, for generic API requests, the method and endpoint.

• Deletes wait for approval by default; you can also block or allow them entirely
• A call that changes more than 10 records needs approval (bulk brake, configurable)
• After 20 deletes in an hour, the MCP server rejects further ones and notifies you by email, Slack or Teams (emergency brake)

Your own rules per tool, source and argument

For a single tool or every tool of a source, you decide whether a call runs without approval, needs approval or is blocked. A rule can apply only to deletes or only to a specific argument, such as object type “tasks” or a budget up to 500. Blocks always win; otherwise the most specific rule applies.

• Platform and organization admins can lock rules. Business unit users then can't remove them and can only add stricter ones.
• The agent knows its rules up front, from the server instructions, every tool description and the revprime__rules tool.
• Tools that are always blocked aren't offered to the agent at all.

Approve where you already work

Per MCP server, you choose what needs confirmation on top of that: nothing (the default), all write tools or selected tools. The approval shows the tool, business unit, server and the full arguments. Once you approve, the action runs with exactly those arguments. If you reject it, nothing happens.

• In the hub on the Approvals page
• Through an approval link, for example to your head of sales or the CRM owner
• In Slack with “Approve & run” and “Reject” buttons when the Slack app is connected, otherwise with a link to the approval page
• In Microsoft Teams as a card that links to the approval page

Limits for every agent

Each MCP server has its own limits, by default 500 calls per hour and 50 write calls per 24 hours. Calls sent in parallel or in a batch can't get past them either. When a limit is hit, the agent gets a clear error and you get a single notification.

Access levels and careful defaults

Every connection is set to Read, Write or Admin. For Google Analytics, Search Console, Tag Manager, LinkedIn and Meta, the hub only requests the permissions of that level when connecting. Campaigns and ads are created as drafts or paused, and most Google Ads changes run as a dry run by default.

Vendor rights in view

A tool being enabled in the Data Hub doesn't mean the vendor allows it for the connected account. Under “Rights at the vendor”, each MCP server lists the tools that are missing there and how to fix it. The agent learns this through revprime__rules and doesn't try those tools in the first place.

• Granted permissions at Google, LinkedIn, HubSpot, Meta and Salesforce, e.g. a missing Marketing Hub
• Roles in LinkedIn and Meta ad accounts and in Search Console properties
• Vendor refusals in the log, including API keys

Undo with one click

If an agent changed something it shouldn't have, you revert it in the activity log. The hub stores no field contents for this, only references such as the record ID and field name. It fetches the old value from the vendor's history and doesn't overwrite anything someone changed since.

• HubSpot: changed fields, created records, notes, tasks and associations
• Campaign status in Google Ads, LinkedIn Ads, Meta Ads, Instantly and HeyReach (except archiving at Meta, which is final there)
• References are kept for 30 days