MCP Servers for Small Business: What they are and what is safe to connect

MCP Servers for Small Business: What They Are and What Is Safe to Connect

An MCP server is a small program that lets an AI assistant such as Claude or ChatGPT read from, and sometimes act in, one of your business tools, for example your accounting software, CRM, or payment account. For a small business the useful question is not “what is MCP” but “which connections are worth making, and with what permissions”. The short answer: start with read-only access to one tool, prefer servers published by the vendor itself, and keep anything that can move money or delete records behind a human click.

This guide is general information, not a security audit or legal advice. Facts about specific servers were checked against the vendors’ own documentation and the MCP specification on 1 November 2026. Vendors change these pages often, so confirm before you connect anything.

Key Takeaways

  • MCP (Model Context Protocol) is an open standard for connecting AI applications to external systems. The official site compares it to a USB-C port for AI.
  • The MCP specification itself says tools “represent arbitrary code execution and must be treated with appropriate caution”, so a connected server is a trust decision, not a plug-in toy.
  • Stripe, HubSpot, Xero and Intuit each publish an MCP server of their own. Their designs differ: some run remotely and sign you in with OAuth, others run on your own computer with API credentials in a config file.
  • Give each connection the smallest permission that does the job. Several of these servers let you restrict scopes or turn off write, update and delete tools.
  • Never connect payroll, bank logins, passwords, or records you are legally required to protect until you understand exactly what the server can do.

What Is an MCP Server, in Plain Terms?

An MCP server is a translator between an AI assistant and one external tool. The assistant asks in plain language (“which invoices are overdue?”), the server turns that into the tool’s own request, and the answer comes back to the chat. The official MCP site describes MCP as “an open-source standard for connecting AI applications to external systems” covering data sources, tools, and workflows.

Three pieces are involved:

  • The host or client is the AI app you chat with, such as Claude Desktop or Claude Code.
  • The server is the connector for one system, such as Stripe or Xero.
  • The tools are the specific actions the server offers, such as “list invoices” or “create invoice”.

The tools list is what you should read as an owner. A server that only offers “list” and “get” tools can look things up. A server that also offers “create”, “update” or “delete” tools can change your records. The same server can be harmless or risky depending on which tools are switched on.

MCP Server vs Plain API vs Zapier or n8n

An API is the doorway every software tool already has for other software. An MCP server wraps that doorway so an AI assistant can use it without custom code. Automation tools such as Zapier and n8n do something different: they run fixed workflows you design in advance (“when a form arrives, create a contact”), while an AI assistant using MCP decides which tool to call in the moment. If no server exists for a tool you use, our guide to creating an MCP server without coding shows the no-code routes.

That difference matters for risk. A fixed workflow does the same thing every time. An AI assistant choosing tools can be steered by text it reads, which is why the permissions you grant matter more. A separate guide compares the three approaches in detail. If you are weighing no-code automation first, our n8n vs Make vs Zapier comparison is the better starting point.

Official, Vendor-Run and Community Servers

Before you connect a server, find out who publishes it. There are three common cases:

  1. Published by the vendor, hosted by the vendor. You connect to a web address the vendor runs and sign in with OAuth. Stripe and HubSpot work this way.
  2. Published by the vendor, run on your computer. You install the vendor’s code locally and give it credentials. Xero’s and Intuit’s servers work this way.
  3. Published by someone else. A community project, a directory listing, or a reseller. These may be fine, but nobody at the vendor stands behind them.

The table below records what each vendor’s own documentation says, as of 1 November 2026.

Two cautions on that table. First, “published by the vendor’s organisation on GitHub” is not the same as a polished, supported product. Read the repository’s own description of what it is for. Second, a server being official does not make it safe to give full access. It only tells you who to hold responsible when something changes.

What Stripe Says About Its Server

Stripe’s documentation describes a server that lets agents “interact with the Stripe API” and search Stripe’s knowledge base. It tells users to “turn on your MCP client’s tool-approval setting so you can review tool calls before they run”. Some tools that move funds between Treasury accounts are marked in the tool list as requiring human confirmation.

Stripe also posts a deadline for key-based setups: Stripe says to migrate by 30 November 2026; starting in December, Stripe MCP stops accepting full-access secret keys and restricted keys without the Agent tag, and expects agent API keys or OAuth instead. If you set this up with an old key, check the notice on Stripe’s page.

What HubSpot Says About Its Server

HubSpot describes a remote server that gives AI tools “read and write access to your HubSpot CRM data”. According to HubSpot’s current integration docs (last modified 6 November 2026), you create an MCP connector in your HubSpot account and connect your AI tool to it using OAuth with PKCE. All actions respect your existing HubSpot user permissions. HubSpot says the server does not allow access to custom Sensitive Data Properties, and that if Sensitive Data is turned on in your account, activity and conversation data is blocked. HubSpot’s landing page says the server is primarily aimed at developers and technical teams, so a non-technical owner may need help with setup. For how HubSpot compares with Pipedrive, Zoho CRM and Salesforce, see our guide to CRM MCP servers for small business.

What Xero and QuickBooks Publish

Xero’s repository describes an MCP server that bridges the protocol to Xero’s API for contacts, the chart of accounts, and invoices, with a longer tool list that also includes payroll for some regions. Its own setup notes recommend starting with a Demo Company and note that scopes can be limited through an environment variable.

Intuit’s repository describes a local server that runs “as a stdio subprocess” on your machine and authenticates to a QuickBooks Online company. Its configuration lets you suppress create, update and delete tools. Both servers keep credentials in a config or environment file on your own computer, so that file is as sensitive as a password.

Our QuickBooks and Xero MCP server guide walks through connecting both safely.

What an AI Connection Can Do Wrong

The MCP specification is blunt about the risk. In its security principles, it says tools “represent arbitrary code execution and must be treated with appropriate caution”, and that descriptions of what a tool does “should be considered untrusted, unless obtained from a trusted server”. It also says AI apps must obtain explicit user consent before invoking any tool, while noting that the protocol itself cannot enforce these principles.

In practice, for a small business that comes down to four risks.

Too much permission. An assistant that can read invoices does not need to delete them. If a server offers broad access and you grant it, any mistake or manipulation uses all of it.

Text that steers the assistant. An email, web page or document the assistant reads can contain instructions. Stripe warns about this by name when it says to take care combining its server with others “to avoid prompt injection attacks”. The more tools connected at once, the more damage one steered request can do. For a worked example with email, see our guide to connecting Claude to Gmail safely.

Local servers on your computer. The MCP security guidance says local servers “may have direct access to the user’s system”, and lists attacks such as a malicious startup command in a client configuration. Only install servers from sources you can verify, and treat copied config snippets as code.

Credentials in plain files. Local servers usually need keys stored in a config file. Anyone or anything that reads that file gets the same access you gave the server.

Our guide to the risks of another agent tool, is OpenClaw safe for business, walks through the same pattern in detail. For a test, permission levels and a checklist to use with any connector, see our guide to MCP security for small business.

Start Safe: The Least-Permission Rules

Use these rules for every connection.

  1. One tool first. Connect a single system, learn what it does, then decide on a second.
  2. Read-only first. Choose read scopes where the vendor offers them. For servers with switches, disable create, update and delete tools until you have a reason to enable them.
  3. Test in a demo or sandbox. Xero recommends a Demo Company and Intuit supports a sandbox. Use them.
  4. Turn on approval prompts. Keep your AI app asking before each tool call, as Stripe advises.
  5. Use a dedicated account where possible. An account made for the connection can be revoked without locking you out.
  6. Keep money movement manual. Payments, payouts and transfers should require a person every time.
  7. Review and revoke. Check connected apps monthly and remove what you do not use.

What to Connect, What to Think Twice About, and What to Avoid

This is a judgement based on the documentation above, not a vendor ruling.

If your business handles health, children’s, or other regulated data, check your legal obligations with a qualified adviser before connecting anything.

What It Costs

The MCP standard itself is open source. What you pay for is everything around it: the AI plan you chat with, whichever plan of each business tool includes API or app access, and the time or help needed for setup. This page does not quote prices, because they change and should be read from each vendor’s own pricing page.

Where to Go Next

The cluster of guides this page anchors will cover choosing servers, connecting accounting and CRM tools, WordPress, security, and cost:

Frequently Asked Questions

Do I need to be technical to use an MCP server?

For hosted servers like Stripe’s and HubSpot’s, setup is mostly signing in with OAuth, though HubSpot says its server is aimed at developers and technical teams. Local servers such as Xero’s and Intuit’s need developer credentials, Node.js, and a config file, so most owners will want help.

Is MCP the same as an API?

No. An API is a tool’s own interface for software. MCP is a standard wrapper that lets AI apps use many tools’ APIs in one consistent way.

Are community MCP servers safe?

Some are fine and some are not. The MCP security guidance warns that local servers can access your system, so install only from a source you can verify and read what the server can do first.

Should I let an AI assistant send payments?

Not without a human approving each payment. Stripe marks its fund-moving tools as requiring human confirmation and advises reviewing tool calls before they run.

Will an MCP connection give the AI company my data?

Anything the assistant reads during a chat is processed by that AI product under its own terms. Check the privacy terms of the AI plan you use before connecting customer data.

Sources

  • Model Context Protocol, “What is the Model Context Protocol (MCP)?” and Specification, version 2026-07-28 (modelcontextprotocol.io)
  • Model Context Protocol, Security Best Practices (modelcontextprotocol.io)
  • Stripe, Model Context Protocol documentation (docs.stripe.com/mcp)
  • HubSpot, HubSpot MCP server (developers.hubspot.com/mcp)
  • Xero, xero-mcp-server README (github.com/XeroAPI/xero-mcp-server)
  • Intuit, quickbooks-online-mcp-server README (github.com/intuit/quickbooks-online-mcp-server)

All retrieved 1 November 2026.