What Is an MCP Server and Does Your Business Need One?
What MCP is, what an MCP server does, which assistants support it today and when a business is better off having one. With primary sources and dates, and no programming tutorial.
Published 10 November 2026
Short answer
An MCP server is the program a business uses to make its product or data available to AI assistants such as Claude, ChatGPT or Gemini through MCP (Model Context Protocol), an open standard that lets an assistant use the data and tools of other systems. Anthropic created MCP in November 2024, and since December 2025 it has been governed by the Agentic AI Foundation, part of the Linux Foundation. A business benefits from one when its customers would want to look up its data or use its product without leaving the assistant; a services website with no data of its own does not. An MCP server does not make AI recommend a business: it is a separate channel, for people who already know it.
What is MCP (Model Context Protocol)
DefinitionMCP (Model Context Protocol) is an open standard that lets an AI assistant, such as Claude or ChatGPT, use tools and data from other systems. The official documentation at modelcontextprotocol.io defines it as an open-source standard for connecting AI applications to external systems.
“Think of MCP like a USB-C port for AI applications,” says the same documentation. Just as USB-C connects devices from different brands through one plug, MCP lets any compatible assistant talk to any compatible system in the same way.
The problem it solves is custom integration. When Anthropic introduced MCP, it explained that every new data source required its own custom implementation. With MCP, a business builds one connection and any assistant that supports the protocol can use it.
What is an MCP server and what does it do
ServerAn MCP server is the program a business puts in front of its data or product so an AI assistant can read it or act on it, following the MCP standard. The assistant acts as the client: it asks what the server offers and uses it when the conversation calls for it.
According to the MCP specification, a server can offer an assistant three things:
- Resources: data and context the assistant reads, such as a catalogue, a customer record or a document.
- Tools: actions the model can run, such as searching a database, creating an order or sending a message.
- Prompts: message templates and workflows that the person chooses to use.
The person using the assistant never sees an “MCP server”. In Claude it appears as a connector, according to Anthropic’s documentation; in ChatGPT, as a plugin, according to OpenAI’s. They switch it on once and then ask in plain language: “find this customer’s pending orders”.
Who created MCP and who governs it today
OriginMCP was created by Anthropic, the company behind Claude, which open-sourced it on 25 November 2024. Since 9 December 2025 it has been governed by the Agentic AI Foundation, a fund within the Linux Foundation co-founded by Anthropic, Block and OpenAI.
A standard that does not depend on a single company matters to anyone building on it. According to Anthropic’s announcement of that date, the Agentic AI Foundation is also supported by Google, Microsoft, Amazon Web Services, Cloudflare and Bloomberg.
The MCP specification is released in dated versions. As of 10 November 2026, the current one on modelcontextprotocol.io is dated 28 July 2026.
How widely MCP is used: servers and downloads
AdoptionMCP is now the standard the major assistants share. According to Anthropic, in December 2025 there were more than 10,000 active public MCP servers, and the official MCP software development kits had more than 97 million monthly downloads across Python and TypeScript.
In the same announcement, Anthropic said MCP had been adopted by ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code, and that AWS, Cloudflare, Google Cloud and Microsoft Azure offered infrastructure to deploy MCP servers.
These figures come from Anthropic, the company that created the protocol, published on 9 December 2025 when it donated MCP to the Linux Foundation. They measure the size of the ecosystem, not how many businesses in a given sector already run an MCP server.
Which assistants support MCP servers in 2026
AssistantsAn MCP server works today in Claude, ChatGPT, Gemini and Microsoft Copilot Studio, but each provider offers it in different places and on different terms. The table summarises each provider’s documentation, checked on 10 November 2026.
| Assistant | What the person sees | Where it works | What the provider says |
|---|---|---|---|
| Claude (Anthropic) | A connector | claude.ai, Claude Desktop, the mobile app, Cowork and Claude Code | The server can be listed in Anthropic’s directory or added as a custom connector |
| ChatGPT (OpenAI) | A plugin | ChatGPT, Codex, deep research, company knowledge and the OpenAI API | For deep research and company knowledge, the server needs two read-only tools: search and fetch |
| Gemini (Google) | A custom connected app | The gemini.google.com web app and the Gemini mobile app | Only for people aged 18 or over in the US, with a personal account and in English |
| Microsoft Copilot Studio | A tool for the agent | The agents a business builds in Copilot Studio | Connects existing knowledge servers and data sources |
For a business, the upshot is that one MCP server reaches several assistants, but not all of its customers equally. In Gemini, for example, someone outside the US cannot add their own server today, according to Google’s help pages.
MCP server vs API vs appearing in AI answers: the differences
ComparisonAn MCP server, an API and visibility in AI answers solve three different problems. The MCP server lets an assistant use your product once someone connects it; the API connects programs to each other; appearing in answers makes the assistant name you to people who do not know you yet.
| Criterion | MCP server | Own API | Appearing in AI answers |
|---|---|---|---|
| What it is for | Letting an assistant use your data or product inside the conversation | Letting another program connect to your product | Getting the assistant to name or recommend your brand |
| Who sets it in motion | The person who connects the server to their assistant | A developer who writes the integration | The assistant, when someone asks about your sector |
| What it takes | A server that follows the MCP specification | Documentation and one integration per case | A site crawlers can read and sources that talk about you |
| Reaches new customers | Only if the person connecting it already knows you | No | Yes: it is where suppliers get discovered |
| Works across assistants | Yes, in those that support MCP | Only where someone integrates it | Yes, each assistant with its own sources |
Example: what an MCP server does for a business
ExampleA real example is HeyTalent, a sourcing tool for recruiters. Its MCP server brings candidate search, contact data and LinkedIn outreach into Claude, and it also works in ChatGPT. The recruiter asks for profiles inside the conversation without opening another tab.
In practice, the recruiter types something like “find me backend engineers in Berlin with Go and Kubernetes, 5 to 10 years of experience, currently at a product company” into Claude. Claude calls the search tool on HeyTalent’s server, gets the profiles back and presents them in its answer. That is the flow shown on HeyTalent’s MCP page for recruiters.
The same pattern works in other sectors. A sales team can connect a company database, a contact enrichment service or its CRM, as HappySales explains in how a sales team uses an MCP server.
An online shop can let the assistant look up its catalogue and stock. When the assistant also buys on the person’s behalf, that becomes agentic commerce, a different decision with its own requirements.
When an MCP server makes sense for your business and when it does not
DecisionAn MCP server makes sense when a business has a product or data its customers would want to look up or use without leaving the assistant. It does not for a services website with no data of its own: there, the assistant can do nothing through a server that it cannot already do by reading the site.
When an MCP server makes sense
- You sell software and your customers already work in Claude or ChatGPT.
- You hold data customers look up often: orders, stock, prices, bookings, reports.
- A recurring task can be asked for in one sentence: search, create, send, book.
- You can maintain the server and its security like any other part of the product.
When an MCP server is not worth it
- You are a services business and your site explains what you do, with no data of your own to look up.
- You expect the server to make AI recommend you to new customers.
- Your customers do not use AI assistants at work.
- Nobody in the business can maintain it or review its permissions.
For a services business, the useful work is making sure assistants can read its site and find sources that talk about it. The starting point is to check whether AI crawlers can read your site.
An MCP server does not make AI recommend you: data from Kulturo and HeyTalent
VisibilityAn MCP server has no say in which brands an assistant names when someone asks for a supplier. The assistant only uses servers the person has connected beforehand, so the server serves people who already know the business. For everyone else, what counts is what the assistant reads about the brand.
That other route already brings in customers. Between 15% and 20% of new customers at Kulturo and HeyTalent come through GEO, according to what they say when they sign. GEO is the work of getting assistants to cite and recommend a brand; how it differs from SEO is covered in what GEO is and how it differs from SEO.
HeyTalent runs both routes at once. It offers its MCP server to people who already use its tool and, separately, gets 2 to 3 demos a week from people who found it by asking ChatGPT, according to HeyTalent’s data. The full figures are in the HeyTalent case study with DoGeo.
What decides whether an assistant names a brand is the site it can read and the third-party sources that talk about it. Why those sources weigh so much is explained in brand mentions in AI search.
What are the risks of connecting an MCP server
RisksThe main risk of an MCP server is that it gives an assistant access to real data and real actions. The MCP specification warns that tools can run arbitrary code and must be treated with caution, and requires the assistant to get the person’s explicit consent before invoking any of them.
What the MCP specification and Gemini’s help pages say about the risks of an MCP server:
- Privacy: under the specification, the assistant must get the person’s explicit consent before exposing their data to a server.
- Security: the specification acknowledges that MCP cannot enforce these principles on its own, and asks implementers for authorisation flows and access controls.
- Confirmation: according to Google, Gemini currently asks for manual confirmation before any write action, while warning that it can make mistakes.
- Responsibility: Google does not control, monitor or secure third-party MCP servers; responsibility lies with whoever connects them.
For anyone commissioning an MCP server, the practical question is what the assistant will be able to do on the customer’s behalf and who reviews those permissions. It is the same conversation already held about any integration that touches customer data.
Keep reading
Let’s talk about your visibility
Book a call with DoGeo. We come to it with a report already prepared on how Google and AI assistants see you today, and on what we would do first.