Until recently, the only way an AI assistant could learn about your business was to read your website or a third-party page about you. The Model Context Protocol (MCP) adds another route: a standard way for AI applications such as Claude and ChatGPT to connect directly to your data and services, look things up and even take actions.
That raises a practical question for marketing and product teams. Should your business publish its own MCP server, so assistants can check your inventory, book appointments or query your documentation in real time? For some companies it is a genuine advantage. For many others it is an expensive way to duplicate a website that already works.
This guide explains MCP in plain English, what a public MCP server lets AI assistants do, the security responsibilities that come with it, and a simple framework for deciding whether it is worth building in 2026.
Key Takeaways
- MCP is an open standard for connecting AI applications to external systems: data sources, tools and workflows.
- An MCP server exposes three kinds of capability: tools (actions), resources (data) and prompts (templates).
- Remote MCP servers run over Streamable HTTP and are typically protected with OAuth; local servers use stdio.
- A public MCP server makes most sense when users need live, structured or account-specific answers that a web page cannot give.
- It is not an SEO ranking factor; it complements, rather than replaces, a crawlable website.
What MCP is, in plain English
The official MCP introduction describes it as an open-source standard for connecting AI applications to external systems. Using MCP, an assistant can reach data sources (files, databases), tools (search engines, calculators) and workflows (specialised prompts) in a consistent way.
“Think of MCP like a USB-C port for AI applications.”
— Model Context Protocol, What is the Model Context Protocol (MCP)?
The analogy works because MCP standardises the connector, not the device. You build one MCP server, and any compatible client can use it. The MCP site lists support in AI assistants such as Claude and ChatGPT and in developer tools such as Visual Studio Code and Cursor.
The moving parts
| Term | What it means | Business example |
|---|---|---|
| Host | The AI application the user is working in | Claude, ChatGPT, an IDE |
| Client | The connector the host creates for each server | Created automatically by the host |
| Server | Your program that exposes capabilities | yourcompany MCP server |
| Tools | Functions the AI can call to take actions | check_availability, create_quote, track_order |
| Resources | Data the AI can read for context | Product catalogue, help articles, price list |
| Prompts | Reusable templates for common tasks | "Compare our plans for a 10-person team" |
| Transport | How messages travel | stdio for local servers; Streamable HTTP for remote servers |
According to the architecture overview, MCP uses JSON-RPC 2.0 messages, clients discover what a server offers through list methods (for example tools/list), and the current protocol version is 2026-07-28, which made the protocol stateless and deprecated some older client features. For remote servers, MCP recommends OAuth to obtain authentication tokens. If your developers built against an older version, budget time to update.
What a public MCP server lets assistants do
Think of your MCP server as a structured, permissioned front door for AI assistants. Depending on what you expose, a user could ask an assistant to:
- Check live stock, prices or delivery dates at a specific store.
- Search your documentation or knowledge base and return exact answers with links.
- Book, reschedule or cancel an appointment after signing in.
- Pull their own account data, such as order history or usage, with their permission.
- Create a quote, support ticket or lead in your CRM.
MCP also has an extension for interactive "MCP Apps" that run inside AI clients, which is closely related to the branded apps covered in our new guide to ChatGPT apps for brands. For browser-based agents that use your website directly, see our existing guide to WebMCP and how to prepare your website for AI agents.
Is it worth building? A decision framework
| Situation | Build an MCP server? | Why |
|---|---|---|
| Mostly static content (brochure site, blog) | Usually no | A crawlable website already serves assistants; consider NLWeb or better content first |
| Large, frequently changing catalogue or inventory | Often yes | Live, structured answers beat stale crawled pages |
| SaaS product with an API | Often yes | Users want assistants to act inside your product |
| Bookings, quotes or account actions | Yes, with strong auth | High-intent tasks; needs OAuth and permission scopes |
| Regulated data (health, finance) | Only with legal and security review | Privacy obligations (e.g. HIPAA in the US) apply to any new access route |
| No developer capacity to maintain it | No | An unmaintained server is a security and reputational risk |
Step-by-step: evaluating and piloting an MCP server
- List the top questions and tasks customers bring to assistants about your category (our new guide to prompt research helps).
- Check whether your website already answers them. If yes, improve those pages first.
- Identify what needs live or private data: stock, pricing, account status, bookings.
- Start read-only. Expose resources and safe lookup tools before any action that changes data.
- Design narrow tools with clear names and descriptions, because the AI chooses tools based on those descriptions.
- Add OAuth and least-privilege scopes before exposing anything account-specific.
- Test with real clients such as Claude and ChatGPT, using the MCP Inspector during development.
- Measure usage: tool calls, errors, completed tasks and assisted conversions.
Security responsibilities
Publishing an MCP server means exposing a new interface to your systems. The official security best practices are detailed, and several points matter to any business:
- No token passthrough. MCP servers must not accept tokens that were not explicitly issued for them.
- Scope minimisation. Start with minimal, low-risk scopes and request more only when a privileged action is attempted.
- Protect state handles. Possession of a cart or workflow ID must never count as authentication.
- Per-client consent if your server proxies a third-party API, to avoid "confused deputy" attacks.
- Logging and monitoring of tool calls, so you can audit what assistants did on users’ behalf.
MCP and AI search visibility
An MCP server is not a ranking factor, and it does not replace being findable. Assistants still discover and recommend brands mainly through web search and the content they can crawl. Microsoft’s NLWeb project shows one way the two worlds meet: every NLWeb instance is also an MCP server, built on the schema.org data your site already publishes (see our new guide to NLWeb). Infrastructure providers are also moving in this direction; Cloudflare’s AI bots blog archive includes multiple 2026 posts on agent traffic and monetisation. Semrush’s guide on preparing your site for AI agents is a useful companion read.
Common mistakes
- Building an MCP server before fixing the website. If assistants cannot read your pages, a server will not fix discovery.
- Exposing too much. Broad scopes and write access on day one increase risk.
- Vague tool descriptions. The model relies on them to choose the right tool.
- No owner. Protocol versions change; someone must maintain the server.
- Ignoring privacy law. Account data accessed via AI clients is still personal data.
- Expecting traffic. MCP usage happens inside assistants; measure tasks and conversions, not sessions.
Related Guides
- NLWeb Explained: Microsoft’s Open Project for Conversational Websites
- Freshness Signals for AI Search: Dates, lastmod and Real Updates
- Syndicated and Duplicate Content in AI Answers: Who Gets the Citation?
Frequently Asked Questions
What is an MCP server?
A program that exposes tools, resources and prompts to AI applications through the Model Context Protocol, so assistants can look up data or take actions in a standard way.
Which AI assistants support MCP?
The official MCP site lists support in AI assistants including Claude and ChatGPT, and in developer tools such as Visual Studio Code and Cursor.
Does an MCP server help SEO?
Not directly. It is not a ranking factor. It can help users complete tasks inside assistants, but discovery still depends largely on crawlable web content.
Is MCP secure?
It can be, if implemented carefully. The official guidance covers OAuth, scope minimisation, token validation and consent. A poorly built server can expose data.
How much does an MCP server cost to build?
It depends on scope. A read-only server over existing APIs is far simpler than one that handles bookings or payments. Request a scoped estimate rather than relying on generic figures.
Conclusion
A public MCP server is a strong move for businesses whose value depends on live data or actions: inventory, bookings, SaaS features and account services. For everyone else, a fast, crawlable, well-structured website remains the priority. If you are weighing the options, our app development and web development teams can scope a pilot, and our 90-day AI search plan shows where it fits. Request a custom quote to discuss your use case.
