trustyourinbox

Connecting AI assistants to your data (MCP)

trustyourinbox publishes an MCP server, so you can connect an AI assistant like Claude or ChatGPT straight to your workspace and ask questions in plain language. Paste one address into your assistant, sign in, and ask things like 'how much spoof mail did my domain get last week'. Here's the setup, what it can do, and how the access stays safe.

What is MCP, and what does this give me?

MCP (the Model Context Protocol) is an open standard that lets an AI assistant talk to an outside service in a structured way. trustyourinbox runs an MCP server, so any MCP-capable client can read your DMARC data and answer questions about it without you exporting anything or writing a query.

Once connected, you can ask in plain language and get an answer built from your real reports. For example:

  • "How much spoof mail did acme.com get last week?"
  • "Which of my domains are failing compliance right now?"
  • "What should I fix next?"
  • "Show me the top sending sources for acme.com over the last 30 days."
  • "List my domains."

The answers come back as clean tables and summaries, computed by the same logic that powers your dashboard. So the numbers your assistant shows you always match what you see in the product.

Step 1: Paste the address into your assistant

The address is https://mcp.trustyourinbox.com. It is also in the app under Settings, then Integrations, then MCP (AI assistants), with a copy button.

claude.ai and Claude desktop

In claude.ai, open Settings, then Connectors, then Add custom connector. Name it, paste the address, and leave the two sign-in choices on the options Claude marks as detected. Click Add, then Connect. The Claude desktop app shows the connectors of your claude.ai account, so it is there too.

ChatGPT

Open Plugins, click +, then Add custom MCP server. Paste the address as the server URL, choose OAuth for authentication, tick I understand and want to continue, and click Create as a plugin. On ChatGPT Business and Enterprise, an admin decides whether members may add custom MCP servers.

Claude Code

claude mcp add --transport http trustyourinbox https://mcp.trustyourinbox.com

Then run /mcp, choose trustyourinbox and sign in.

Other assistants

Any assistant that can sign in to a remote MCP server works the same way: add a remote (HTTP) MCP server with the address, and it asks you to sign in.

Step 2: Sign in and choose what it may do

Your assistant opens our sign-in page. It leads with the one thing we can check, the website the app comes from (for example claude.ai wants to connect), and shows the name the app gives itself smaller, as its own description.

Choose the workspace it connects to. It can always read your DMARC data there; tick Allow changes to let it make the changes you could make on the dashboard, and you are asked to confirm it's you. Click Allow and you're back in your assistant. There is nothing to copy and nothing to renew: the assistant keeps itself signed in. Every owner of the account gets an email when an app is connected.

What it can and can't do

A connected assistant can always read: look at your data and summarize it. It makes changes only if you ticked Allow changes, and then only a few safe actions on your behalf (below). Either way it cannot touch your billing, your members, or another workspace.

It's also scoped to one workspace: the one you chose when you signed in. It can never see another customer's data, and a domain name that isn't in your workspace simply comes back as "not found". The assistant acts as you, so if your access is removed, it stops working.

The read tools answer questions about:

  • your domains, their status, and overall workspace health and alignment,
  • mail failing DMARC for a domain, and a workspace-wide threat overview,
  • who is sending as a domain (known senders vs unknown sources),
  • whether your failures are forwarding or real, broken down by authentication path,
  • how each receiver (Gmail, Outlook, Yahoo) sees your mail,
  • your DKIM selectors and key sizes, reverse-DNS coverage, and BIMI logo setup,
  • compliance against the Google, Microsoft, BIMI, and TLS mandates, your open action items, and what to fix next,
  • and your reports and recent changes. There is also a public toolbox that checks any domain or explains a pasted email header.

If you allowed changes, the assistant can also snooze or dismiss an action item, re-check a domain's DNS, label an unknown sender as a known vendor, and stage a one-click DNS fix (which still goes through the same 5-minute hold, email notice, and read back as the app, and a change to an existing record can be undone), including canceling a staged fix. It can also mark a lookalike domain as known. Every change is recorded in your audit log with the app's name.

Browsing your workspace and autocomplete

Beyond asking questions, the connection exposes your data as browsable resources. Clients that support it let you attach a view directly as context, like your workspace health, your action inbox, or a single domain's posture, without having to phrase a question first. There is a set of workspace views (health, compliance, action items, threats, and domains) and a set of per-domain views (overview, protocol status, senders, subdomains, recommendation, access review, and lookalikes).

Clients that support argument autocomplete will also suggest your real domain names as you type, so you pick from a list instead of guessing the exact spelling. All of this stays scoped to your workspace in exactly the same way: you only ever see and complete your own domains.

Disconnecting an app

Open Settings, then Integrations, then MCP (AI assistants). Under Connected apps you see every app connected to the workspace, who connected it, whether it can make changes, and when it was last used. Disconnect ends its access at once. Anyone who can edit workspace data (owners, admins, and members, not viewers) can connect an app and disconnect any app; anyone can disconnect an app they connected.

Every change an assistant makes is recorded in your audit log and per-domain history the same way a change you make by hand is, labeled with the app's name so you can always see when an action came through an assistant rather than a person.

For scripts and tools that can't sign in

A script, a scheduled job or a tool that cannot open a sign-in page uses a token instead. Create one under For scripts and tools that can't sign in on the same settings page, copy it (you only see it once), and send it with the same address as an Authorization: Bearer tyi_mcp_... header. Tokens are read-only unless you allow write actions, expire, and can be revoked any time; every owner is emailed when one is created. Use a token too for what connected apps can't do: add webhook endpoints, alert recipients and invites.

Treat a token like a password. We store only a one-way hash of it, so nobody (including us) can read it back. If one is lost or shared by mistake, click Revoke next to it and it stops working on the next request.

Troubleshooting

The assistant shows two trustyourinbox connectors. You have two connections, often an older one made with a token. The Claude apps also load the MCP servers set up in Claude Code on the same computer. Keep one and remove the other.

It answers about the wrong workspace. An app reaches the one workspace you chose when you signed in. Disconnect it and connect it again, choosing the other one.

"This app was connected read-only." You left Allow changes unticked. Disconnect it and connect it again with the box ticked.

It can't find my domain. Use the exact domain name as it appears under "list my domains". A domain that isn't monitored in this workspace (or that belongs to a different workspace) returns "not found" by design.

A token is refused as "unauthorized". It was mistyped, has expired, or was revoked. Make sure the header is exactly Authorization: Bearer tyi_mcp_... with a single space and the whole token. If in doubt, create a fresh one.

Still stuck? Open the help bubble in the dashboard and send us a note. We read every one.

Ask your DMARC data anything.

Connect Claude, Cursor, or any MCP client to your workspace and ask the way you would ask a colleague. It reads your real reports and answers with clean tables, the same numbers you see in the app. Included on every plan.

Was this page helpful?

Keep reading.

Start with one domain. See it with your own data.

Free for one domain, no card. Add one DNS record and the first reports usually arrive within a day.

  1. Monitoring
  2. Quarantine
  3. Reject