MB Connect

What it can do

Nine tools, and this is the complete list. Four work without an account; five require you to connect your MB account first.

Permissions and OAuth

Connecting uses OAuth 2.1 with PKCE. Your assistant never sees your MB password — you sign in on our own page, approve a specific list of permissions, and the assistant receives a token limited to exactly those.

Permissions are deliberately narrow and separate. Reading a company profile is a different permission from searching employees, which is different again from creating a document. There is no admin or all-access permission in the system at all. Any permission that touches company data is additionally bound to the single company you select on the consent screen — a token issued for one company cannot read another, even when the same account owns both.

We also refuse to offer permissions we cannot honour. Capabilities that are reserved for future features are not shown on the consent screen and cannot be granted, so a permission you approve is always one that does something.

How your data is handled

Only what a specific tool call needs is sent, at the moment the assistant makes that call. Reading an employee sends that employee’s identifier and returns their record; drafting a document sends the field values for that document. Nothing is sent in the background, and there is no bulk export.

Your resume drafts are not involved at all. They live in your browser’s local storage, and MB Connect has no access to them.

Worth being explicit about: your assistant provider sees the request and the response, because that is how the assistant reads the result. If you ask ChatGPT to draft an offer letter, the field values and the resulting document pass through OpenAI; the same is true of Anthropic when you use Claude. That is inherent to using an assistant, not something particular to us — but it is your data going to a third party, so it belongs in a decision you make knowingly. Our privacy policy sets out each data flow separately.

On our side, connector activity is logged as metadata only — which tool ran, whether it succeeded, how long it took, and a request ID. Document contents, employee names, salaries, and tokens are never written to those logs, and that is enforced by an automated test, not just by policy.

Disconnecting and revoking

Two independent ways, and either one is enough:

• Remove the connector inside ChatGPT or Claude. The client stops using its token.
• Revoke the grant from our side. This invalidates the token itself, so it stops working even if a client still holds a copy. This is the stronger action — use it if you are unsure who still has access.

Access tokens are short-lived, and refresh tokens rotate on every use — an old refresh token cannot be replayed after a newer one has been issued. A revoked grant cannot be quietly resurrected.

Troubleshooting

Support

Email support@mbresumebuilder.com. Every response from the server carries a request ID — including it lets us find the exact call in our logs without you sending us any document content or employee data.

Frequently asked questions

What is MB Connect?

MB Connect is our Model Context Protocol (MCP) server. It lets an AI assistant you already use — such as ChatGPT or Claude — work with MB HR Documents on your behalf: look up which document types exist, check what fields a document needs, draft one, and render it to PDF or DOCX. It is a single server at https://mbresumebuilder.com/mcp, shared by every client, so both assistants get identical tools and identical permissions.

Which assistants can connect?

Any client that speaks MCP over Streamable HTTP. We test against ChatGPT (Developer Mode custom connectors), Claude (custom connectors), and the official MCP Inspector. Because there is one server rather than a per-vendor integration, a client we have not tested that implements the specification correctly will also work.

What can the connector actually do?

Nine tools. Four need no account at all — listing document types, checking a document's required fields, drafting a document from values you supply, and rendering a draft to PDF or DOCX. Five require you to connect your MB account with OAuth: reading a company profile, searching employees, reading one employee, listing issued documents, and saving a document as a draft. That is the complete list; there is nothing hidden behind it.

Can it send, sign, or revoke documents?

No. Sending a document to a recipient, requesting a signature, and revoking an issued document are deliberately not exposed to any AI client. Those actions exist only in the MB web app, where a human performs them directly. Drafting and creating are separated from sending and signing on purpose: an assistant can prepare work for you, but it cannot deliver a document to an employee or put anything in front of a signatory.

What permissions am I granting?

You approve specific, named scopes on our own consent screen before anything is shared — never a blanket 'allow this app' toggle. Scopes are narrow and separate: reading a company profile is a different permission from searching employees, which is different again from creating a document. There is no admin or all-access scope in the system. Any scope touching company data is also bound to the one company you pick on that screen, so a token issued for one company cannot read another, even when the same account owns both.

What data leaves my browser when I use the connector?

Only what the specific tool call needs, at the moment the assistant makes it. Reading an employee sends that employee's identifier and returns their record. Drafting a document sends the field values for that document. Your resume drafts are never involved — they live in your browser's local storage and the connector has no access to them. Your assistant provider (OpenAI or Anthropic) will also see the request and response as part of your conversation with it, because that is how the assistant reads the result.

How do I disconnect or revoke access?

Two independent ways, and either is sufficient. Remove the connector inside ChatGPT or Claude, which stops the client using its token. Or revoke the grant from our side, which invalidates the token itself so it stops working even if the client still holds it. Revoking from our side is the stronger action; use it if you are unsure whether a client still has a copy. Access tokens are short-lived and refresh tokens rotate on every use, so a revoked grant cannot be quietly resurrected.

Do I need a paid plan?

The four anonymous tools are free and need no account. The account-scoped tools need a signed-in MB account. Rendering a document to a real PDF or DOCX consumes the same Tool Pass or HR Pro entitlement the website's own generator pages consume — the connector does not create a cheaper path around the paywall, and it does not create a more expensive one either.

It says my tool needs a scope I thought I granted. What now?

Disconnect and reconnect the connector, and check the consent screen lists the permission the tool needs. This usually means the original connection requested a narrower set of scopes than the tool you are now calling requires. If it persists, contact support with the request ID from the error — every response carries one, and it lets us find the exact call without you sending us any document content.