MCP for business: connect your tools to AI assistants

What the Model Context Protocol is, what changed in the July 2026 spec, and how to connect your CRM and business tools to ChatGPT or Claude safely.

Crest Design
AI assistant connected by lines to business app icons

The Model Context Protocol (MCP) is an open standard that lets AI assistants such as ChatGPT and Claude connect to your business tools: your CRM, help desk, database, store or internal systems. Instead of copying data into a chat window, the assistant can look things up and, if you allow it, take actions through a defined set of tools. For most businesses the sensible starting point is to use the connectors your existing software already offers, then build a custom MCP server only for the systems nobody else covers.

This guide explains what MCP is, what changed with the July 2026 specification, and how to roll it out without creating a security problem, at the time of writing (November 2026).

What MCP actually does

MCP defines a common way for an AI application (the client) to talk to a service that exposes capabilities (the server). A server describes its tools (actions such as "create a ticket" or "look up an order"), resources (data the assistant can read) and prompts (reusable templates). The assistant decides when to call a tool, and the server does the work with its own credentials and rules.

The benefit is that you build an integration once and it can work with any MCP-compatible assistant, rather than writing a separate plugin for each one. Anthropic created MCP and, in December 2025, donated it to the Agentic AI Foundation, a fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. That vendor-neutral home is one reason support is now broad.

What changed in the 2026-07-28 specification

MCP's maintainers released a new version of the specification on 28 July 2026, which the official announcement describes as moving to a stateless core. The changes that matter for businesses are:

  • Stateless requests: the old start-up handshake and session IDs are gone. Each request carries what the server needs, so any instance behind an ordinary load balancer can answer it. That makes hosting MCP servers simpler and more reliable.
  • Easier routing and caching: requests now include standard headers that gateways can use to route and meter traffic, and tool lists carry cache hints.
  • Tighter authorisation: stricter checks on who issued credentials, and a move from dynamic client registration towards Client ID Metadata Documents.
  • Deprecations with a long runway: features such as Sampling, Roots and Logging are deprecated, with at least twelve months before removal under the new lifecycle policy.

The TypeScript, Python, Go and C# SDKs support the new version, with Rust in beta. If a developer built you an MCP server earlier, ask whether it relies on session IDs, because that part will need migrating.

Where you can use MCP today

Assistant How custom MCP servers are added Notes
Claude Custom connectors using remote MCP Available on all plans, with Free limited to one custom connector. On Team and Enterprise, owners add connectors and each member connects their own account (Claude help centre).
ChatGPT Add a custom MCP server from the Plugins page Needs a public HTTPS endpoint or OpenAI's Secure MCP Tunnel. Workspace policies decide who can add servers (OpenAI docs).
Automation platforms Agents that call MCP servers as tools For example, n8n Agents can connect an MCP server and exclude specific tools.

Plan details and screens change often, so check the vendor's current documentation before rolling out to a team.

A practical rollout plan

1. Start with one job, not one system

Pick a task people already do by hand, such as "summarise this customer's open tickets and recent orders before a call" or "draft a reply using our returns policy". That tells you which data the assistant needs and which actions, if any, it should be allowed to take.

2. Use official connectors first

Many SaaS vendors now publish their own MCP servers or connectors. These are maintained by people who know the product and usually support proper sign-in. Use them before building anything.

3. Build a custom server for the gaps

Custom servers make sense for in-house databases, legacy systems or a bespoke booking engine. Keep the tool list small and specific: "get order status by order number" is safer and more reliable than "run any SQL query". Write clear tool descriptions, because the model relies on them to decide what to call. Our post on Node.js integrations between stores and CRMs covers the kind of API groundwork that often sits behind a server like this.

4. Start read-only, then add actions

Give the assistant read access first and watch how it is used. When you add write actions, require confirmation for anything that changes records, sends messages or spends money.

5. Test before you share

The MCP Inspector tool (run with npx @modelcontextprotocol/inspector@latest) lets developers check a server's tools and responses before connecting it to an assistant.

Security: the part not to skip

An MCP connection gives an AI model a way to reach real systems, so treat it like any other integration with access to customer data. This is general guidance, not a security audit.

  • Least privilege: use dedicated service accounts or per-user OAuth with the narrowest scopes possible. Never connect an admin key.
  • Prompt injection: content the assistant reads (emails, tickets, web pages) can contain instructions designed to trick it. Researchers at Palo Alto Networks' Unit 42 have documented MCP-related attack paths. Confirmation steps and narrow tools limit the damage.
  • Trust your servers: Anthropic's help centre notes that it does not verify third-party remote MCP servers. Only connect servers from vendors you trust or code you control.
  • Logging: record which tools were called, by whom and with what inputs, so you can review behaviour and investigate problems.
  • Data protection: check where data flows and whether it is covered by your agreements with the AI vendor, especially for personal data.
Small, specific tools with confirmation on anything that writes are the safest way to start.

Frequently asked questions

Do I need a developer to use MCP?

Not to use existing connectors, which are usually a few clicks to add. Building, hosting and securing a custom server does need development work.

Is MCP the same as an API?

No. MCP usually sits on top of your existing APIs, describing them in a way AI assistants understand and can call safely.

Will an MCP server work with every assistant?

That is the aim of the standard, and remote servers using Streamable HTTP work with the major assistants. Each client supports features slightly differently, so test with the ones your team actually uses.

Need help connecting your tools to AI?

We design and build MCP servers, API integrations and the guardrails around them as part of our AI and automation services, and our Node.js team handles the hosting and back-end work. For example, we built a React front end and Node.js API for a startup incubator. If you are new to automation, five workflow automations to set up first is a good primer. To plan your first connector, contact us.

Sources: MCP blog: 2026-07-28 specification, MCP specification changelog, Linux Foundation: Agentic AI Foundation, Claude help centre: custom connectors, OpenAI: Connect from ChatGPT, Unit 42: MCP attack vectors

(over to you)

Planning a build like this?

Tell us what you're building and we'll come back with ideas and next steps.