What an MCP connector actually does for your business
Not protocol internals — what changes about your day when an assistant can reach a system you already use.
Quick answer: what is an MCP connector?
An MCP connector is a standard way to give an AI assistant or agent access to a real system you already use, with permissions you control. MCP — the Model Context Protocol — is an open standard, so one connector works across every assistant that supports it instead of needing a separate integration built for each one. In practice it is the difference between an assistant that can tell you how to file a receipt and one that files it, in your own account, because you approved the connection.
Most writing about MCP is aimed at people implementing it. This page is for the other audience — the founder, bookkeeper or operator who keeps seeing “connectors” in a settings menu and wants to know what one buys them. The short version: a connector turns an assistant from something that talks about your work into something that can do a specific, bounded piece of it.
What a connector actually does — a real accounting-firm example
The most useful thing we can show is not a diagram. It is what happened when an accounting firm connected ExpenseBot to the AI agent it was already running.
We did not build a custom integration. That is the whole point of the example. The firm connected ExpenseBot over MCP — the open standard that lets AI assistants securely use external tools and data — and wired it to the agent they already run. No project, no API contract negotiated between two vendors, no waiting on a roadmap.
The division of labour ended up like this:
“The MCP connection itself is working well, and it gives us a practical way to combine ExpenseBot’s receipt capture with Viktor’s access to Zoho and our accounting rules.”
One distinction matters more than any other on this page, so it is worth stating plainly: ExpenseBot has no direct Zoho API sync. The firm’s agent holds the Zoho access, not us. Blurring that line would make the story sound better and make it false. What the connector did was make the firm’s own agent able to reach receipt data it previously had no route to.
Nor does the agent “do the books.” It verifies and reconciles under the firm’s rules, and the firm reviews the result. The deeper walkthrough of that arrangement — including the review-first instructions they gave the agent — is written up at using ExpenseBot as the evidence source for a reconciliation agent.
Now the generalisation, which is the actual payload. The same ExpenseBot connection works with ChatGPT, Claude, Gemini Spark, desktop assistants and other MCP-compatible agents. You connect it to the agent you already use, approve access to your ExpenseBot account, and decide what role it plays. One person submits a receipt photo from Claude. Another asks ChatGPT for a client expense report. This firm combined receipt capture, AI-assisted reconciliation and Zoho Books — without forcing any one product to do everything.
Kinds of MCP connections
There are three shapes, and which one you use is decided by the assistant, not by you. The full setup instructions for each live on the ExpenseBot MCP page — specifically its connection chooser, which is the canonical version. This is the one-paragraph orientation.
| Kind | Typical clients | What you do |
|---|---|---|
| Directory install | ChatGPT, Claude.ai | Find it in the connector directory and click Connect, then sign in. No URL, no config file, no token. |
| Custom remote (OAuth) | Gemini Spark, custom business agents | Add the remote MCP address as a custom app or integration and complete browser sign-in. |
| Local stdio + token | Claude Desktop, Cursor, other local clients | Add a small entry to the client’s config file and store a named, revocable token in its own secret settings. |
The thing worth internalising is that authentication is not the same across clients. A directory install never involves a token; a local client does. Treating them as interchangeable is how tokens end up pasted somewhere they should never be.
MCP connectors worth knowing about
There is no single official league table, and any list published on a vendor’s site goes stale. The two places that stay current are Claude’s connector directory and the official MCP Registry. Treat those as authoritative and this as a map of the territory — the categories below are what connectors are actually for, checked 2026-08-10.
| Category | What a connector there lets an assistant do |
|---|---|
| Developer tools | Source control, issue trackers and error monitoring — so an assistant can read a stack trace or open a pull request without you switching windows. |
| Databases and files | Query a database in plain language, or let an assistant read a document out of cloud storage instead of you pasting it in. |
| Productivity and comms | Notes, task boards and chat — the assistant can look up what was decided rather than asking you to find it. |
| Design | Read a design file's structure and turn it into markup or a spec. |
| Web infrastructure | Deploy, inspect logs, and manage DNS or edge config from the conversation. |
| Payments and cards | Look up a charge, a payout or a subscription without opening the billing dashboard. |
| Accounting and spend platforms | Read transactions and, on some platforms, code or approve them. Capability varies widely — check what each one actually writes. |
| Receipts and expenses | ExpenseBot: submit a receipt photo from the chat, search what you spent, scan Gmail for receipts you never forwarded, and build a report — into a Google Sheet you own. |
The honest caveat when you are comparing them: “has a connector” and “can do the thing you want” are different claims. Some connectors only read. Some write but gate most of their surface behind an administrator account. Some cannot accept a file from the conversation at all. Before you commit to one, ask what the specific action you care about does — not whether the logo appears in a directory.
Security: what a connector can and cannot do
A connector is an access grant, so it deserves the same scrutiny as any other one. The model ExpenseBot uses, and the questions worth asking of anything else you connect:
- Permissions travel with the user. A connector acts as you and can only reach what you can reach. It cannot see another person’s data.
- Remote connections use browser OAuth — you approve on a consent screen, and no credential is copied around.
- Local connections use a named, revocable token stored in the client’s own secret settings. Never paste a token into a chat message or an email.
- Verify read-only before you let anything write. Ask for a summary first — “show my top five expense categories this month, don’t change anything” — and confirm it returns something you recognise.
- Your data stays where it already lives. ExpenseBot’s primary expense records are in a Google Sheet in your own Drive; the connector reads and writes there rather than moving your records somewhere new.
- Revocation should be one step. Disconnect in the assistant, or revoke the named token, and access ends.
For a connector attached to an accounting workflow, one more rule is worth borrowing from the firm above: start it read-only, have it propose matches and exceptions, and keep every write behind explicit approval until you have watched it be right on your own books.
If you want the setup steps rather than the concepts, they are all on the MCP page. If you want to see what to actually ask once it is connected, the prompt guide is a list of things that work. Assistant-specific walkthroughs live at ExpenseBot in ChatGPT and the Claude connector.
One practical note before you start: whether a given connector is available to you can depend on your assistant’s plan, your region, and your workspace administrator’s settings. Those are the assistant vendor’s gates, not ours — the Connect button in the directory is the current source of truth. ExpenseBot’s connector is free for ExpenseBot users, with a 60-day free trial and no credit card.
Connect ExpenseBot to the assistant you already use
Submit a receipt from the chat, ask what you spent, scan Gmail for receipts you never forwarded — into a Google Sheet you own. $10/month, 60-day free trial, no credit card.
Start free — no credit cardFrequently asked questions
What is an MCP connector?
An MCP connector is a standard way to give an AI assistant or agent access to a real system you already use. MCP — the Model Context Protocol — is an open standard, so one connector works across assistants that support it rather than needing a separate integration for each. You approve the connection yourself, and the connector can only do the specific actions it publishes. It is the difference between an assistant that can describe how to file a receipt and one that actually files it.
Is an MCP connector the same thing as an MCP server?
They are two words for the same component seen from different sides. "MCP server" is the technical term for the software exposing the tools; "connector" is what the assistant's interface usually calls it once you add it. If you are choosing something to use rather than building one, the practical difference is nil.
Do I need to be a developer to use an MCP connector?
No, for the common path. In ChatGPT and Claude you add ExpenseBot from the connector directory and sign in — no config file, no URL to paste, no token. Only local desktop clients require editing a configuration file, and that route is documented separately for people who want it.
Can I use ExpenseBot's connector with an agent that isn't Claude or ChatGPT?
Yes. The same remote connection works with Gemini Spark as a custom Connected App, with desktop clients, and with custom business agents that support remote MCP. Authentication differs by client — some use browser sign-in, some need a revocable token in their own credential store — so follow the setup route for the specific client you are connecting.
What can an MCP connector see in my ExpenseBot account?
Only what you authorise, and only as you. Permissions travel with the signed-in user, so a connector cannot reach another person's data. Your primary expense records stay in your own Google Drive; ExpenseBot stores the limited account, authorization, audit and duplicate-prevention metadata needed to run and secure the service.
How do I revoke a connector's access?
Disconnect the connector in the assistant, or revoke the connection from ExpenseBot. For local clients that use a named token, revoke that token in AI Assistant Tokens and it stops working immediately. Access is designed to be reversible without touching your data.
Does an MCP connector do anything when I'm not in the chat?
The connector itself acts when an assistant calls it, so it is not a background job. ExpenseBot's own nightly Gmail receipt scan runs on its own schedule regardless, and an assistant can trigger a scan on demand rather than waiting for it.