Agents that are allowed to work in your CRM, ad accounts and outbound.

The Model Context Protocol is an open standard that lets AI agents like Claude work with other systems. The Data Hub gives each agent its own MCP endpoint. Through it, the agent sees exactly the tools you've enabled.

No. Anyone who only needs to connect tools, such as the HubSpot admin or someone on your ads team, gets a personal connect link. It's valid for 1 to 30 days, and you can revoke it at any time. Through the link, that person connects only the tools you picked, with instructions for each one, and can't see any other data. If you want, you can also invite them to the Data Hub; they'll then see only their own business unit.

The Data Hub doesn't keep a copy of your data. It stores configuration, encrypted credentials and the activity log. Content is fetched from the vendor on every call and only passed on to the agent. For approvals, the arguments and result of that single action are kept, encrypted. On top of that, it keeps a few details without any content: references for undo (record IDs, field names and the previous campaign status, 30 days), monthly usage counters (counts only) and, for the rights check, the connected account's role per account.

Only if you enable write tools for its MCP server. Even then, deletes and calls that touch more than 10 records need approval by default, and an emergency brake halts runs of deletes. With your own rules, you can pre-approve routine work, require approval or block actions, including based on arguments. There are also limits per day and per hour. Campaigns are created as drafts or paused, and most Google Ads changes run as dry runs by default. If the agent applies Google's optimization recommendations, though, they take effect right away.

Any MCP client that supports Streamable HTTP. That includes chat assistants like Claude, ChatGPT, Gemini Enterprise, Microsoft 365 Copilot, Grok, Perplexity and Mistral Le Chat; agents like OpenClaw, Hermes Agent, Manus, Notion custom agents and HubSpot Breeze; automation with n8n, Make, Zapier and Copilot Studio; and developer tools like Claude Code, Codex, opencode, Gemini CLI, GitHub Copilot, Cursor, Windsurf, Zed and Kiro, including when they run inside herdr. Most log in with OAuth and only need the MCP server URL. A token in the header works too. For platforms without self-registration, you set up an OAuth client by hand.

For sales-led B2B SaaS and AI companies. It fits if your pipeline lives in HubSpot or Salesforce and you use agents in marketing, sales and RevOps. A typical setup has one business unit per team, region or product line. The Data Hub isn't built for online stores: there are no connectors for shop systems, and the field definitions and reports revolve around deals, opportunities and pipeline.

The connectors return more than raw data. With every result, the agent gets the definitions of the fields it contains, for example that the deal stage in HubSpot is an internal ID, that open deals can have a close date in the past, or that Salesforce keeps the weighted amount as ExpectedRevenue. On top of that come reports across sources (ad spend against deals and pipeline per campaign and channel), ready-made playbooks like the weekly pipeline report, and templates for typical agents.

Yes. You can set up each team, region or product line as its own business unit with its own connections and agents. Anyone in marketing, sales or your SDR team who should look in can get a login to exactly that business unit. New users join by invitation only.

No. Each tool works with the permissions granted when the account was connected, often with fewer, because the connection's access level and allowed accounts narrow them further. If the account lacks a permission in the tool itself, the Data Hub can't add it.