Connecting AI assistants to your business tools with MCP
Short answer: The Model Context Protocol (MCP) is an open standard that lets an AI application such as Claude or ChatGPT talk to outside systems (your CRM, calendar, files or database) through a small program called an MCP server. The assistant can then read data and, if you allow it, take actions. Start with one read-only connection to a low-risk tool, check exactly which permissions you grant, and keep approval switched on for anything that writes, sends or deletes.
Founders hear "just connect it to your CRM" and imagine one switch. In practice there are a few moving parts, and each of them is a place where you choose how much the assistant can see and do. Once you understand those parts, deciding what to connect becomes much less of a gamble.
What MCP is
The project's own introduction describes MCP as an open-source standard for connecting AI applications to external systems: data sources such as files and databases, tools such as search, and workflows such as prepared prompts. Its analogy is a USB-C port for AI applications. Instead of every AI app building a separate integration for every tool, a tool exposes one MCP server and any MCP-capable app can use it.
The specification is the formal version. Messages use JSON-RPC 2.0, and three roles talk to each other.
Hosts, clients and servers
- Host: the AI application you use, such as a desktop chat app or a coding tool.
- Client: a connector inside the host. The architecture page explains that the host creates one client per server, each with its own connection.
- Server: the program that exposes your system. A CRM vendor might publish one, or a developer might build one for an in-house tool.
What a server can offer
According to the spec, a server can offer three kinds of feature:
| Feature | What it is | Business example |
|---|---|---|
| Resources | Context and data for the user or model | Reading a customer record or a price list |
| Prompts | Templated messages and workflows | A "summarise this account" routine |
| Tools | Functions the model can run | Creating a task, updating a deal stage, sending an email |
Tools are where the risk sits. Reading a record and changing one are very different permissions, even when they come through the same server.
Local or remote
The architecture page notes that local servers typically use the STDIO transport and serve a single client, while remote servers use Streamable HTTP and typically serve many clients. In plain terms: a local server runs on one person's computer; a remote server is hosted on the internet and can be shared.
Claude's custom connector guide adds a detail that matters for a company network: when you add a remote connector, Claude connects from Anthropic's cloud, so the server must be reachable from the public internet. A server behind your VPN won't connect unless you allowlist their IP ranges.
Hosted connectors vs your own server
Most founders meet MCP through connectors already listed inside an AI app. Those are quick to set up and maintained by someone else. A self-built server is for tools that have no connector, or when you want tight control over which actions exist at all.
A useful rule: if a vendor's official connector exists, start there. If you need a custom server, have it expose only the few functions you actually need. A server with three narrow tools ("look up customer by email", "list open orders", "add a note") is far easier to trust than one with full API access.
Your first safe project, step by step
Here is a cautious sequence for a small firm, using a hypothetical example: a five-person agency wants an assistant to answer "what's the status of client X?" from its project tool.
- Pick a read-only question. Status lookups, summaries and searches. No sending, no editing.
- Choose a low-risk data source. A project board is a better first test than your mailbox or accounting system.
- Use the vendor's connector or a narrow server. Check who built and hosts it. Claude's guide says to connect only to servers built and hosted by organisations you trust.
- Read the permission screen properly. During sign-in, the connector asks for scopes. Claude's guidance is to limit scopes where possible and deny access if requested permissions seem unnecessary.
- Keep approvals on. The spec says hosts must obtain explicit user consent before invoking any tool. In OpenAI's API, the MCP tool guide recommends using
require_approvalandallowed_toolsso sensitive actions need sign-off and only chosen tools are loaded. - Test with awkward inputs. Ask about a client that doesn't exist, two clients with similar names, an archived project. Note where the answers are wrong.
- Decide on write access only after a few weeks. If you add it, start with one harmless action, such as adding a comment, and keep approval on.
- Know how to switch it off. Claude's guide says permissions can be revoked by disconnecting the connector in its settings or in the third-party service's security settings.
On personal data: decide in advance which systems an assistant may read and check that this fits how you already handle customer data.
Risk table
| Risk | What it looks like | What reduces it |
|---|---|---|
| Prompt injection | Text inside a document, email or tool response tells the assistant to do something you didn't ask | Approvals on write actions; avoid connecting tools that read untrusted content and send data out in the same session |
| Over-broad access | A connector asks for full account access when you need one list | Deny unnecessary scopes; narrow custom servers |
| Untrusted server | A server you found online that you can't vouch for | Use official connectors or servers you or your developer control; OpenAI warns a malicious server can exfiltrate data from anything that enters the model's context |
| Misleading tool descriptions | A tool says it is read-only but isn't | The spec treats tool descriptions as untrusted unless the server is trusted |
| Token mishandling in custom servers | A server passes your login token straight to another API | The MCP security best practices call this "token passthrough" and say the authorisation spec forbids it |
FAQ
Do I need a developer to use MCP?
Not for connectors already listed inside your AI app; adding those is a settings task. You need a developer to build or host a custom server for a tool with no connector.
Is MCP only for Claude?
No. The MCP introduction lists support across Claude, ChatGPT, Visual Studio Code, Cursor and other clients. What each app allows, and on which plan, differs, so check the app's own help page.
Can an assistant delete data through MCP?
Only if the server exposes a tool that can and you grant the permission. That is why narrow servers and approvals matter more than any single setting.
Should I connect my email first?
It's usually a poor first choice. A mailbox holds a lot of sensitive information and receives text from strangers, which is exactly where prompt injection starts.
When you know which question you want an assistant to answer from your own tools, we can scope a narrow, read-only connection around it.
Sources
- https://modelcontextprotocol.io/docs/getting-started/intro
- https://modelcontextprotocol.io/specification/2026-07-28
- https://modelcontextprotocol.io/docs/learn/architecture
- https://modelcontextprotocol.io/specification/2025-06-18/basic/security_best_practices
- https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp
- https://developers.openai.com/api/docs/guides/tools-connectors-mcp