What is the MCP server protocol? Start with a picture. A Large Language Model (LLM) is a brain that understands language, reasons and proposes solutions: it is the engine behind the chatbots and AI assistants we use every day. For a long time, though, this brain had no arms. It could think, but it could not act on the systems where a company keeps its data.
MCP servers are the arms. If the language model is the brain, the MCP server is the part that lets it open a database, query a CRM, read a document or send a command to a business application. This step is what separates a chatbot that answers from an AI agent that works, the same difference between chatbots and AI agents people talk about so much. In this guide we look at how the Model Context Protocol works, where it stands today and how to use it in a company without taking on needless risks.
Table of Contents
- What is MCP Server?
- What is the MCP Server protocol used for?
- How MCP Server Works
- The state of the MCP protocol today
- How the user experience of chatbots is changing
- MCP Server Use Cases
- MCP server security: risks and good practices
- How to Implement MCP Server in Your Company
- Conclusions
- FAQs about MCP Server
What is MCP Server?
An MCP server is a program that gives an AI model access to tools, data and prompts through the Model Context Protocol, the open standard created at Anthropic and now managed by the Linux Foundation.
The protocol was launched by Anthropic in November 2024 to connect AI assistants to enterprise data sources in a simple and secure way. The aim is to have the models we use in Claude, ChatGPT or GitHub Copilot work on real, current data, instead of only on what they learned in training.
You can think of the protocol as a universal interpreter between the chatbot and the rest of the digital world: databases, APIs, CRMs, knowledge bases, ERPs, ticketing systems. Before MCP, every link between an LLM and an external source had to be custom-built, and each vendor did it its own way. With a shared standard, a connector written once works with every application that speaks MCP.
The protocol was designed to:
- standardize the links between AI and applications;
- handle requests that depend on the context of the conversation;
- cut the work of building and maintaining integrations.
The architecture has three parts:
- MCP Host → the AI application the user works in (for example Claude, ChatGPT or Visual Studio Code), where the model runs;
- MCP Client → the connector inside the host that keeps the link with a single MCP server;
- MCP Server → the program that exposes tools (actions), resources (data) and prompts to the model, connected to external data sources and applications.
What is the MCP Server protocol used for?
An MCP server lets the language model talk to the environment it works in. With the MCP protocol a chatbot can:
- retrieve current information from external sources;
- take specific actions, such as opening a ticket, sending an email or querying a database;
- answer based on the company's real data;
- do all of this within the permissions it has been given.
An example. Without MCP, a chatbot can explain the returns policy, but it does not know the status of the order of the person writing, whether a product is in stock or what that customer bought last year. With MCP the bot reads this data and, if allowed, changes it: it becomes a digital operator. This is the step you need to create AI agents that close a case instead of passing it to a colleague.
How MCP Server Works

The language model can reason and write answers, but on its own it cannot touch company systems. The MCP server is the bridge: when the model needs data or an action, it goes through it. The cycle looks like this:
- the server receives the model's request;
- it finds the right source (CRM, ERP, database, internal APIs);
- it retrieves or updates the information;
- it returns current context to the model, which builds the answer on it.
Compared with RAG (Retrieval-Augmented Generation), which retrieves chunks of documents that were indexed in advance, MCP works on live systems: it reads the current status of an order and can also change it. The two techniques often live in the same agent. The benefit of the standard is that each system is connected once, with one set of rules for authentication and permissions, and the connector keeps working if you switch model or application tomorrow.
The MCP protocol architecture in detail
To see how it works technically, follow a real request:
- A user asks the chatbot: "What is the status of my order #12345?"
- The LLM reads the request and works out that it needs the order system;
- The MCP client forwards the request to the right MCP server;
- The MCP server queries the orders database with the expected credentials;
- the data goes back to the LLM with the context it needs;
- The LLM answers: "Your order #12345 has shipped and should arrive tomorrow."
For the user it is an ordinary conversation that takes a few seconds.
The state of the MCP protocol today
In two years MCP has gone from an Anthropic project to a shared industry standard. In December 2025 Anthropic donated it to the Agentic AI Foundation (AAIF), a Linux Foundation fund that now also hosts other agent standards such as A2A and AGENTS.md. The Linux Foundation announcement of 14 September 2026 says so, and it is the same release that launched MCPA, the first official certification for people who design and build with MCP.
The announcement also gives a sense of how widely the protocol is used. The official Tier 1 SDKs are approaching half a billion downloads a month, and the TypeScript and Python SDKs have each passed 1 billion total downloads. MCP tool calls from ChatGPT users in August 2026 were 98 times their January level. MCP now works in ChatGPT, Claude, Visual Studio Code with GitHub Copilot and many other clients: a server written once can be used by all of them.
The 2026-07-28 specification
On 28 July 2026 the new version of the MCP specification was released. These are the main changes listed in the official changelog:
- stateless core: protocol-level sessions and the initial handshake are gone. Each request carries the client's version and capabilities, so a remote server can be run like any other HTTP service;
- Multi Round-Trip Requests, which replace the requests the server used to send to the client (sampling, elicitation, roots);
- cacheable results for tool and resource lists;
- stricter authorization, with token issuer validation and credentials bound to their issuer;
- formal extensions framework, which brings MCP Apps (interactive interfaces shown inside the chat) and Tasks (long-running operations that continue in the background).
For a company, the practical result is that servers are easier to scale and to run in production. Teams with connectors built on earlier versions have time to migrate: Roots, Sampling and Logging are deprecated but keep working for at least 12 months.
How the user experience of chatbots is changing
People writing to the chatbot notice the difference at once, because the bot stops saying "I don't have access to that information". With the MCP protocol the experience becomes:
- More personalized: the bot reads CRM data and past interactions, and adapts the tone and content of its answer.
- More responsive: the bot fetches the information it needs by itself, instead of sending the user to another channel or to an agent.
- More reliable: answers are based on data read at that moment from company systems, which cuts "hallucinations" about orders, prices or stock.
- More controlled: the protocol includes authentication and authorization, so the bot only sees the data and actions it has been granted.
In practice the chatbot knows what it is talking about, because it reads the right data at the moment it is needed.
MCP Server Use Cases

The protocol shows its value when you look at it inside a real process. Here are five use cases where connecting an LLM to company systems changes the way people work.
1. Customer Care
A chatbot connected via MCP to the CRM and the order system answers customers with their own data, updates tickets, alerts operators when something goes wrong and proposes a fix. This is the basis of AI agents for customer care, which close repetitive requests on their own and leave the hard cases to people.
Practical example: a customer asks about a delayed shipment. The chatbot queries the logistics system via MCP, gets the parcel status, finds the cause of the delay and offers a refund or a priority reshipment, within the rules set by the company.
2. Internal operational support
AI assistants can help employees by consulting ERPs, inventories and internal knowledge bases. An AI agent can guide staff through order management, stock checks or activity planning, with fewer errors and less time lost looking for information spread across three different systems.
Practical example: a warehouse operator asks the chatbot whether a product is available. The bot queries the ERP and, if the item is out of stock at the main site, lists the other warehouses that have it.
3. Sales and marketing data-driven
With access to customer, campaign and product sales data, an agent can suggest cross-selling and up-selling offers, prepare reports on request and start targeted communications. The sales team gets in one question what used to take a CRM export and a spreadsheet.
Practical example: an AI agent looks at a customer's recent purchases in the CRM history and proposes complementary products, to raise the average basket value.
4. Compliance and document management
In legal and regulatory work, the protocol lets AI assistants consult internal policies, regulations and contracts. They can answer complex questions, check whether a process is compliant and reduce errors in sensitive tasks, always under the supervision of the people responsible.
Practical example: a legal office uses a chatbot connected via MCP to the document archive. The bot retrieves contract clauses, checks whether a process complies with GDPR and drafts standard documents in seconds.
5. Continuous innovation and integration of new tools
With MCP, adding a data source or an application does not mean redesigning the system. The new tool is connected once and becomes available to the agents already running, which add skills one connector at a time.
Practical example: a company adopts a new analytics tool and builds a dedicated MCP server for it. From then on every chatbot and AI agent already in operation can use it, without rewriting existing integrations.
MCP server security: risks and good practices
Connecting a language model to company systems gives it access to real data and real actions, and that opens new risks. Stacklok's State of MCP in Software 2026 report names security as the top obstacle to adoption, cited by 64% of the technical leaders surveyed. The same report finds that nearly half of software companies already use MCP in production, but only 19% have rolled it out broadly.
Two risks get most of the attention. Tool poisoning happens when a tool description contains hidden instructions: the model reads them as part of its context and may take actions the user never asked for. Indirect prompt injection comes through the data instead: a ticket, an email or a web page read by the connector can contain text written on purpose to manipulate the agent. Both come down to the same problem, an agent that trusts everything that enters its context.
A few rules, applied consistently, reduce the risk:
- use only servers built in-house or from known vendors, and read tool descriptions before you enable them;
- give each server the minimum permissions: a customer care agent that reads order status should not be able to cancel orders;
- ask the user or an operator to confirm actions that change data or cost money;
- manage access through your company identity provider. The Enterprise-Managed Authorization extension, stable since June 2026, enables connectors for employees through single sign-on, without a separate OAuth consent for each app. Okta is the first supported identity provider;
- log tool calls, so you have a trail for debugging and audits.
The roadmap published in August 2026 lists agent identity and enterprise-ready security among the project's five priorities. The maintainers' stated goal is for servers to recognise agents through existing standards, instead of pasted API keys and long-lived tokens.
How to Implement MCP Server in Your Company
The Model Context Protocol gives AI agents standard, controlled access to company systems. To get results, though, the project has to start from a precise goal: answering customers faster, taking load off a team, having sales data one question away. These are the steps we recommend.
Step 1: Map company information sources
Every organization uses a set of tools: CRM, ERP, databases, knowledge bases, ticketing systems, intranets. Adopting MCP is a chance to put data flows in order, see which sources matter most and decide which information AI assistants should be able to reach and which they should not.
Step 2: Involve the right people
An MCP project connects business and technology, so IT, operations, marketing and customer care should work on it together. People in the teams know which questions come in every day and which repetitive actions are worth automating; IT knows which systems can be exposed and with which permissions.
Step 3: Choose reliable partners and solutions
Platforms like Crafter.ai let you connect conversational agents and MCP tools without building everything from scratch. A specialized partner lowers the risk of a poor integration and supports the company from project design to measuring results.
Step 4: Define success metrics
Before you start, set clear KPIs: customer service response times, first-contact resolution rate, operating costs, NPS. You also need an estimate of volumes: our AI agent traffic calculator helps you work out how many conversations the agent will handle. With these numbers you can measure the return on the project and see where to act.
Step 5: Train the team and iterate
An MCP integration does not end with the first release. The team needs training on AI tools, user feedback has to be collected and configurations reviewed regularly: tools that are rarely used get removed, missing ones get added.
MCP in Crafter.ai
Crafter.ai uses the MCP protocol in two directions. With MCP Tools the bot connects to external tools: you can build a custom tool visually, without writing code, or connect an existing HTTP MCP server. For each tool you decide whether the agent can see it, and you can restrict its use with tags. The MCP Tools tutorial shows how to connect the chatbot to a database or an API.
In the other direction, Crafter.ai has its own authoring MCP server: with your account you can read and change the bot configuration in plain language from Claude Code, Claude Desktop, GitHub Copilot, OpenCode or claude.ai. And if you are interested in the web side, the page on WebMCP and AI agents explains how a website can offer its content as tools to agents browsing from the browser.
Conclusions
MCP servers give language models the access to company systems they used to lack. With an open standard, managed by the Linux Foundation and supported by ChatGPT, Claude and Copilot, chatbots become an active part of business processes: they read data, take actions and answer based on what is actually happening in the company.
For a company, this means moving from chatbots that talk to AI agents connected to its own tools, provided permissions, confirmations and call logging are in place from day one.
If you are considering how to adopt the MCP protocol in your organization, Crafter.ai offers an AI agent platform ready to connect to the main enterprise systems. Write to the team here for a personalized consultation.
FAQs about MCP Server
What is an MCP Server?
It is a program that connects an AI model to external applications, APIs and databases using the Model Context Protocol, the open standard launched by Anthropic in 2024. It acts as a bridge between artificial intelligence and a company's information systems.
Why is it important for chatbots?
Because it lets the bot read current data, take concrete actions such as opening a ticket or changing an order, and answer with the company's real information instead of only what it learned in training.
Is MCP Server compatible with all AI models?
MCP depends on the application, not on the model. Today ChatGPT, Claude, Visual Studio Code with GitHub Copilot and many other clients support it natively, and the official TypeScript and Python SDKs have each passed 1 billion total downloads.
What changes with the July 2026 MCP specification?
Version 2026-07-28 makes the protocol stateless: no more sessions or initial handshake, so remote servers scale like ordinary web services. It also adds stricter authorization and the MCP Apps and Tasks extensions.
What is the difference between MCP and an API?
An API is the service that does the work, for example the order management system. MCP is the standard way an AI agent discovers which actions are available and calls them. Often an MCP server simply presents existing APIs to the model.
Is it difficult to implement?
It depends on how many systems you want to connect and how they are built. With a platform that already supports MCP, like Crafter.ai, you connect an existing server or build a tool without code; the real work is deciding on data and permissions.
What are the main advantages of the MCP protocol over traditional integrations?
With MCP each system is connected once and the connector works with every compatible application. That means less code to write and maintain, and the freedom to change AI model without redoing your integrations.
Is the MCP protocol secure?
The protocol includes authentication and authorization, but security depends on how you implement it. Use trusted MCP servers, give each tool the minimum permissions, ask for confirmation before sensitive actions and log every call.





