Connecting an AI Assistant to WarrantyHub with the MCP Server

Connecting an AI Assistant to WarrantyHub with the MCP Server

Connecting an AI Assistant to WarrantyHub with the MCP Server

Who this is for: WarrantyHub company administrators (full admins) who want to let an AI assistant such as Claude or ChatGPT work with their WarrantyHub data.

1. What the MCP Server is

MCP stands for Model Context Protocol. It is a common standard that lets an AI assistant connect directly and securely to a business system — the same way you might connect a spreadsheet app to your bank account. WarrantyHub's MCP Server is the doorway that lets an AI assistant you already use work with your WarrantyHub information.

Once you connect, you can ask an assistant plain-English questions instead of clicking through screens, and you can ask it to do things too. It can look up a contract or a claim, explain how your plans and pricing are set up, run and save your reports — and it can file and update claims, import contracts, add service technicians, build out policy pricing and coverage, and set up your workflow automation and claim statuses.

A few things that are true of every connection:

  • Your data only. A connection can only reach the company (and, if you have more than one, the single program) it was approved for.
  • Your permissions only. The assistant acts as the person who approved it, with that person's permissions.
  • Nothing changes without your say-so. Every tool that writes data previews first. The assistant shows you exactly what would change and writes nothing until you confirm. See "How changes are made safely" below.
  • Approved by a human. Every connection has to be approved by a company administrator on a consent screen before it works, and the consent screen spells out whether the connection can only read, or can also change data.

2. Where to find it in WarrantyHub

Go to Settings → MCP Access (the tile is labelled MCP Access, "Manage connected AI agents"). The page has three sections:

  • MCP Access — your Server URL, your Client ID, your Agent limit, and the Connect instructions for each type of AI client.
  • Connected agents — the assistants currently connected in this program, and a Revoke button for each.
  • Registered clients — apps that registered themselves against your server URL, and a Disable button for each.

3. If you don't see MCP Access

The page is hidden unless two things are both true:

  1. The MCP Server feature has been installed for your company. It is not self-service — only WarrantyHub support can turn it on.
  2. You are signed in as a full administrator. The tile is deliberately restricted, because approving and managing AI connections is the same level of trust as being a full admin yourself.

If you are a full admin and the tile still isn't there, contact WarrantyHub support and ask them to install the MCP Server feature for your company. If the tile is there for a colleague but not for you, ask your own company administrator to raise your account to full administrator.

4. How to connect

You never type your password into the AI assistant. You give it your Server URL, and it sends you back to WarrantyHub to sign in and approve the connection.

Step by step

  1. Open Settings → MCP Access.
  2. Under Connect instructions, expand the section for the assistant you use: Claude Code, Claude (Desktop & Web), ChatGPT, Codex CLI, Cursor, VS Code, or Other MCP clients.
  3. Use the copy button next to the value or command shown there. The instructions already have your company's real Server URL filled in — you don't have to assemble anything yourself.
    • Claude (Desktop & Web): in Claude, go to Settings → Connectors → Add custom connector and paste the Server URL.
    • ChatGPT: go to Settings → Connectors → Create (an eligible ChatGPT plan is required) and paste the Server URL.
    • Cursor: use the Add to Cursor one-click button, or paste the configuration block shown.
    • Claude Code / Codex CLI: copy the command shown and run it in a terminal.
    • Other MCP clients: any client that supports remote (HTTP) MCP servers just needs the Server URL. The panel also shows a bridge command for older clients that only support local servers.
  4. The assistant sends you to a WarrantyHub sign-in page. Sign in as yourself.
  5. You land on the approval screen. Read it before clicking:
    • It names the application asking, and says it wants to act "on your behalf as [your name] @ [your company]".
    • A banner tells you whether the application is running on your own computer, is a verified vendor application, or is unverified. If it says unverified, only continue if you recognise the web address shown.
    • This connection will be able to: lists exactly what it gets, in plain terms — for example reading warranty contracts and customers; filing new claims; updating claims and their claim jobs; importing contracts in bulk; cancelling contracts; running and exporting reports; saving report definitions and schedules; searching, creating and editing service technicians; and reading or changing company configuration. Read this list carefully: it is where you decide how much the assistant can do.
    • If your company has more than one program, pick the Program the connection is limited to.
  6. Click Allow (or Deny to stop). The assistant is now connected, and it appears on the Connected agents list.

Two useful extras on the page:

  • Copy setup guide (Markdown) copies every client's instructions as one document you can paste into an email or chat for your team.
  • Agent limit caps how many assistants can be connected at once across all your programs (10 by default). You can change it and click Save. If someone hits the cap, the approval screen tells them so and asks them to revoke an existing connection or raise the limit.

5. What you can do once connected

Ask questions

A connected assistant can look up individual contracts and claims, find service technicians, explain how your company is configured, and work with your reports — including generating a CSV export and handing you a download link for larger data pulls.

  • "Which WarrantyHub company and program am I connected to, and what am I allowed to do?"
  • "Look up the contract for Jane Miller on Oak Street — is it still active, what's covered, and have there been any claims?"
  • "Pull up claim 10432 and summarise what's happened on it so far."
  • "Walk me through how our Premier plan group is set up and priced."
  • "What automated emails do we send when a claim is approved?"
  • "How many claims did we open last month, broken down by status?"
  • "Export all active contracts that expire in the next 90 days to a CSV I can download."

Get work done

The assistant can also make changes for you — always previewing first (see below).

  • "File a claim on contract A-10233: the compressor failed on the 12th. Show me what you'll create before you do it."
  • "Move claim 10432 to Approved and add an internal note explaining why."
  • "Here's a spreadsheet of 40 new contracts from our dealer — import them, and tell me about any rows that won't go through."
  • "Add these three service technicians."
  • "Set up pricing for our new Platinum plan from this SKU list, then show me the coverage each policy ends up with."
  • "Create a workflow rule that emails the homeowner when their claim is marked Complete, and save the email template it uses."
  • "Save that claims-by-status report and email it to me and Dana every Monday morning."

How changes are made safely

Every tool that writes data runs in preview mode by default. The assistant validates your request, reports exactly what would change — new records, old value to new value, and any rows that would fail — and writes nothing. It only applies the change after you confirm the preview. Two more safeguards worth knowing:

  • Bulk work is row-by-row. On imports of contracts or service technicians, one bad row fails only that row; the rest still process, and re-sending a row that's already correct does nothing.
  • Destructive actions say so. Cancelling a contract cancels every coverage period on it and can't be undone through the assistant, so it always previews and asks for an explicit confirmation first.

A few practical notes:

  • Large data pulls are exported to a CSV file in the background rather than listed in the chat. The assistant gives you a download link, valid for 7 days (it can generate a new one if it lapses).
  • Exports are rate-limited. If you ask for several in a row, the assistant will be told to wait rather than failing silently.
  • The assistant can only see and do what your own WarrantyHub account can — and only what was approved on the consent screen.

6. We update the MCP Server regularly — refresh your connection

The MCP Server is actively developed. We add new capabilities, improve existing ones, and refine how the assistant is guided to use them. Your AI assistant, however, learns what's available at the moment it connects — so a long-running connection can be working from an out-of-date picture of what WarrantyHub can do.

Every month or so, and any time you read about a new capability, refresh your connection:

  1. In your AI client, disconnect and reconnect the WarrantyHub connector (or simply restart the client — most clients re-check on start-up).
  2. If you're asked to sign in and approve again, read the This connection will be able to: list before clicking Allow. New capabilities sometimes need permissions your original approval didn't include, and the assistant can't use them until you grant them.
  3. Ask the assistant "what can you do in WarrantyHub now?" to see the refreshed list.

Refreshing is safe and takes a few seconds. It doesn't affect anyone else's connections, and it doesn't change your Server URL.

7. Revoking access

Because a connection can change data as well as read it, revoking promptly matters. There are two levels, and it's worth knowing which one you need.

Revoke one connection

On Settings → MCP Access, find the assistant in the Connected agents table (it shows the agent name, who connected it, its permissions, and when it was last active) and click Revoke. Confirm in the dialog. That agent's very next request fails, and it cannot reconnect without someone approving a fresh consent screen.

When to use it:

  • Someone left the company. Revoke every connection listed under their name — the connection carries their permissions, including the ability to change data, and keeps working after they stop logging in.
  • You don't recognise a connection. Revoke first, ask afterwards.
  • A tool is no longer in use. Housekeeping, and it frees a slot against your agent limit.

Rotate the connect URL (the "change all the locks" option)

Click Rotate connect URL at the top right of the MCP Access card and confirm. This mints a brand-new Server URL and disconnects every connected agent, in every program. Everyone has to reconnect using the new connect instructions.

When to use it:

  • Your connection details leaked — the Server URL was pasted into a public channel, a ticket, a shared document, or a repository.
  • A shared device or account was compromised, and you can't be certain which connections are affected.
  • Routine credential rotation, if your organisation's security policy requires it on a schedule.

Because rotation disrupts everyone, prefer revoking individual connections when you know exactly which one is the problem.

8. Disabling registered clients and ending sessions

Some AI apps can't be handed a Client ID and instead register themselves automatically against your Server URL the first time they connect. Those registrations are listed under Registered clients, with the app name, the web address it sends sign-in results back to, how many live connections it has across all programs, and when it registered.

To end one, click Disable and confirm. Any connections belonging to that app stop working immediately.

When to use it:

  • Suspicious activity. If a registration you don't recognise appears — especially one whose "Redirects to" address you don't know — disable it. That cuts off every session it holds at once, without waiting to identify each connection individually.
  • Ending access immediately. Disabling is instant; there is no waiting for a session to time out.
  • Freeing up connections. Your company can hold up to 25 active registrations and, by default, 10 connected agents at once. Disabling registrations you no longer use keeps room for the ones you do. When you're at the registration limit, WarrantyHub automatically disables the oldest registration that has no connections to admit a new one.

Disabled registrations are hidden from the list by default; tick Show disabled to see the history.

Need help?

If the MCP Access page isn't visible, if a connection won't complete, or if you think a connection has been misused, contact WarrantyHub support. If you suspect misuse, use Rotate connect URL first — it takes effect immediately — and then get in touch.