# AI agents for small businesses: connectors, MCP and what they can do

> An AI agent is an assistant that acts: it uses tools, reads your data and works through several steps. A connector (built on MCP) is what gives it access to a program such as Backbone. Custom connectors work in Claude on every plan, including Free with one connector (as of 08/10/2026); in ChatGPT they depend on your plan and workspace settings.

Source: https://back-bone.io/en/blog/ki-agenten-kleinunternehmen  
Published: 2026-10-21 · Updated: 2026-10-22  
Category: Guides

An AI agent is an assistant that acts: it uses tools, reads your data and works through several steps to reach a goal. A connector, built on the open MCP standard, is what gives it access to a program such as Backbone. Custom connectors work in Claude on every plan, including Free with one connector (as of 08/10/2026); in ChatGPT they depend on your plan and workspace settings.

**TL;DR · Key facts**

- **Agent:** an assistant with tools. It reads, decides the next step, acts and shows you the result.
- **Connector / MCP:** the plug between the AI app and your software. MCP is an open standard, a bit like USB-C for AI.
- **Today in a back office:** log time, draft invoices from hours, find overdue invoices, prepare reminders, find missing receipts, answer questions about your numbers.
- **Plans (as of 08/10/2026):** Claude custom connectors on Free (one), Pro, Max, Team and Enterprise. ChatGPT: check your plan and workspace, details below.
- **Safety:** you approve each sign-in, the agent works with your role, and sensitive steps need your confirmation.

## What makes an assistant an agent?

A chat assistant answers from what you type. An **agent** also has:

1. **Tools.** It can do something: look up a record, create an entry, draft a document.
2. **Your data.** It reads from your real systems instead of pasted copies.
3. **Several steps.** It plans, calls a tool, looks at the result and goes on, until the job is done or it needs you.

Example: "Invoice last month's hours for Muster GmbH." A chat assistant explains how. An agent finds the hours, builds the invoice draft and tells you what it did. You check it and decide.

There's no hard line between assistant and agent. Many small businesses start with one connected assistant that does one thing at a time on request, and stop copying and pasting. For the bigger picture, see [Agentic back office](https://back-bone.io/en/blog/ki-backoffice).

## What is a connector, and what is MCP?

**MCP** (Model Context Protocol) is an open standard that lets AI apps such as ChatGPT or Claude connect to data sources and tools. Its own introduction compares it to a USB-C port: one standard plug instead of a cable for every device.

A **connector** is such a connection in practice: you add a program's MCP address to your AI app, sign in to the program and approve. From then on the AI app can use what the program offers. A connector you add yourself by address is called a **custom connector**. More in [What is MCP?](https://back-bone.io/en/blog/was-ist-mcp).

## What can an agent do in your back office today?

It depends on what the connected program offers. With a back-office tool, typical jobs are:

- **Time:** "Log 2 hours yesterday on the website relaunch."
- **Invoices and quotes:** draft an invoice from the hours of a month, prepare a quote.
- **Payment reminders:** list overdue invoices, oldest first, and prepare reminders in your tone. Background: Payment reminders.
- **Receipts:** find payments with a missing receipt. See [AI for bookkeeping prep](https://back-bone.io/en/blog/ki-buchhaltung-vorbereiten).
- **Questions:** "Who still owes me money?" or "How many hours went into this client in September?"

What stays with you: deciding, approving and anything legal. An agent drafts and prepares; you check amounts, VAT treatment (Germany and Austria differ, for example in the details an invoice must contain) and what is sent.

Pair an agent with a [skill](https://back-bone.io/en/blog/ki-skills-backoffice) and it follows your own way of working.

## Which plans support custom connectors?

Status as of 08/10/2026, from the vendors' own pages. Plans and menus change often, so the vendor page wins.

| AI app | Custom connectors (remote MCP) | What to know |
| --- | --- | --- |
| ChatGPT | Depends on plan and workspace | OpenAI's current developer pages do not list plans. When we last read them (04/10/2026), developer mode was listed for Pro, Plus, Business, Enterprise and Education on the web. In workspaces, admins may control it. OpenAI is also renaming parts of the feature, with "plugin" now used on its developer pages. Check your own settings. |
| Claude | Free, Pro, Max, Team, Enterprise | Free allows one custom connector. On Team, only owners add one; on Enterprise, owners and roles with the right permission. Each member then connects individually. |

Backbone's AI connector is a custom connector for ChatGPT and Claude today, and also for Perplexity, Cursor, VS Code and Make; Gemini only as Gemini Enterprise. Grok is not covered here. How to set it up: [Connect Backbone to Claude](https://back-bone.io/en/blog/backbone-claude-verbinden) or [Connect Backbone to ChatGPT](https://back-bone.io/en/blog/backbone-chatgpt-verbinden).

## Tips for getting started: using AI agents safely

A few habits get a lot out of an agent and keep you in control. Most of these tips come from the vendors' own help pages:

- **Connect servers you know.** The AI vendor does not verify custom connectors, so the choice is yours. You get Backbone's address straight from your own settings.
- **Read along before you approve.** A server or a document can contain text meant to steer the AI (prompt injection). A quick look at what the AI proposes is enough before you approve.
- **Grant only the permissions the task needs.** An agent can read, change or delete data in connected services, depending on what you allow (Anthropic describes this for Claude). Switch off tools you do not need and approve write actions deliberately.
- **Let each person sign in on their own.** On team plans, a shared key gives everyone the same access. With sign-in per person, everyone works with their own role.
- **Start with a plan that includes a DPA when client data is involved (DE and AT).** What the agent reads from your tools goes to the AI vendor. For client data, the GDPR (Art. 28) requires a data processing agreement with the vendor, usually only offered on business plans. The AI Act also asks you to make sure your team knows how to use AI (Art. 4): plan a short introduction. More in [Agentic back office](https://back-bone.io/en/blog/ki-backoffice).

Look for these points when you choose a tool:

- **Sign-in with approval instead of a token to paste.** You choose which workspaces the AI may use and can revoke it at any time.
- **Your role applies.** The agent sees only what you see.
- **Confirmation for sensitive steps.** Sending and deleting need your OK; bank details stay in the app.

The rule of thumb for any agent: let the AI draft, you approve. Sending, paying and deleting stay with a person.

> My goal is to optimise and automate back-office tasks so that we founders can focus on what matters most.
>
> Automating back-office work with AI agents and AI integrations is the future – that's why I built Backbone!

## How Backbone does it

Your back office in the chat. In the app, or in a chat with ChatGPT, Claude and many more.

1. **Track your time** One sentence in the chat. The hours are in your week.
2. **Bill your hours** What you tracked goes on the invoice.
3. **Create the invoice** Say what goes on it. You approve.
4. **Who's overdue?** One question, every open invoice.
5. **Send the reminder** You approve. Only then does it go out.
6. **Receipts matched automatically** Upload the statement. Clear matches assign themselves.
7. **Permissions per person** Everyone gets what they need. Their AI gets no more.
8. **Review, then approve** The AI drafts. You check it and issue it.

## Frequently asked questions

### What is an AI agent in simple words?

An assistant that uses tools to get a job done. It reads your data, takes steps and shows you what it did.

### Do I need to code to use a connector?

No. You copy an address, add it in your AI app, sign in and approve.

### Does a connector cost extra?

The Backbone connector is part of every plan. Your AI app's own plan and costs depend on your contract with OpenAI or Anthropic.

### Can the agent send an invoice without asking me?

Not in Backbone. It drafts, and it asks before it sends.

### Is MCP only for ChatGPT and Claude?

No. It is an open standard that many tools support. Backbone's guides cover ChatGPT and Claude.

### Is my data safe with an agent?

The agent only sees what your role allows, and you can revoke access at any time. What the AI vendor does with chat content depends on its terms and your plan. For client data, check for a data processing agreement (Art. 28 GDPR) and the training setting.

**A back office for your AI**

Connect Backbone to ChatGPT or Claude and let an agent handle hours, invoices and reminders. 7-day trial, card required.

- [Try for 7 days](https://back-bone.io/en/sign-up)

---

General information, not legal or tax advice and no review of your case. Details on providers and laws can change. For your case, follow your tax adviser.

Checked on 08/10/2026

Sources: Model Context Protocol, official introduction (modelcontextprotocol.io, read 08/10/2026) · Anthropic Help Center, Get started with custom connectors using remote MCP (read 08/10/2026) · OpenAI developer documentation on developer mode and on connecting a custom MCP server (read 08/10/2026) · [Art. 28 GDPR](https://eur-lex.europa.eu/eli/reg/2016/679/oj) · Art. 4 AI Act