What Is an MCP?

MCP gives compatible AI applications a shared way to discover and use server capabilities. This guide explains the client, server and permission boundaries, what documented platform support means, and when a direct integration may be the simpler choice.

Written by MCP Scraper
Published · Updated ·11 min read
A client connects through an MCP boundary to a server exposing declared tools and a source connection.
Editorial schematic; illustrates the article concept rather than a verified product result.
01

MCP Defined

MCP standardizes access to tools, resources and prompts through compatible client-server interfaces. Server implementation, setup and permissions still matter.

Quick take

A shared protocol still needs a working integration.

MCP defines how compatible clients and servers exchange capability descriptions and requests. It does not guarantee that every client supports every feature, or remove server setup, permissions and approval controls. Check the intended connection before relying on it.

What is an MCP in AI?

MCP, the Model Context Protocol, is an open client-server protocol for connecting AI applications to capabilities supplied by external servers. Those capabilities include tools, resources and prompts. A host application connects a compatible client to a server and controls which capabilities are available to the model. The shared protocol can reduce repeated integration work, but it does not remove server implementation, configuration, permissions or user oversight.

Reference: modelcontextprotocol.io

What does the MCP do?

MCP lets a compatible client discover and use capabilities exposed by a server. For tools, tools/list returns available tool definitions and input schemas; tools/call requests a tool operation. Resources and prompts have separate discovery and retrieval operations. A model may select a tool, while the host application controls access and any required approval. In a web research workflow, your AI app is the host. Its MCP client sends the approved tool request to the connected server. The server runs the search or extraction and returns the result through the client. Follow the MCP setup walkthrough and look for extract_url in your client’s tool list. In our November 7 API-key test, discovery listed that tool and the completed extraction returned the heading “Understanding MCP servers.” Check your returned URL and heading against the official MCP server guide. In a separate recorded MCP Scraper example, the search snippet introduced a comparison, while page extraction found a qualification about how the underlying index affects result freshness.

Reference: modelcontextprotocol.io

What is MCP in simple terms?

MCP gives compatible AI applications a common way to use capabilities exposed by a server. A tool server describes callable operations and their input schemas; the client can request an operation and receive its result. Setup still requires matching transport, supported features, configuration and access controls. One server implementation does not guarantee every client can use every feature.

Reference: modelcontextprotocol.io

Is MCP a tool or framework?

MCP is a protocol, rather than an agent framework or a particular tool. It specifies messages and behavior that client and server implementations follow. A framework may embed an MCP client, and a server may expose tools that call an API, query data or perform another operation. Your implementation choice is the client, server or library that meets your task; the MCP specification defines the contract between them.

Reference: modelcontextprotocol.io

What problems does MCP solve?

MCP addresses repeated integration work by giving compatible clients and servers a shared communication contract. Instead of designing a different client-facing interface for every connection, a server can expose defined tools, resources or prompts through MCP. That standardization does not make every client-server combination work automatically. The server still needs its underlying integration, and the client must support the required features, transport and access controls.

Reference: modelcontextprotocol.io

What is Anthropic?

Anthropic is the company behind Claude and the organization that introduced MCP. On December 9, 2025, the MCP project announced Anthropic’s donation of MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation. The announcement says MCP’s maintainer structure remains in place and the project retains autonomy over technical direction and day-to-day operations. Foundation stewardship is therefore different from the Linux Foundation dictating the protocol’s design; the announcement does not establish why each vendor adopted MCP.

Reference: blog.modelcontextprotocol.io

02

MCP vs Everything

MCP vs REST, HTTP, Zapier, Copilot, and LLMs — the comparisons that actually determine whether MCP belongs in your stack.

Key distinction

Choose an interface for a real client requirement.

Start with the operation the client must perform. MCP can provide a shared tool interface; a direct API may already meet the task. Compare the implementation and maintenance work before adding another interface.

What is an MCP vs API?

MCP is a standardized client-server interface for exposing tools, resources and prompts. An API is the broader idea of a software interface; REST is one way to design one. An MCP server may call an existing API, access a database or run other logic. Choosing MCP therefore specifies the client-facing contract, not the backend architecture.

Reference: modelcontextprotocol.io

Why MCP instead of REST API?

Choose MCP when the intended client can use its tool-discovery and invocation interface, or when reusing that interface across clients provides a concrete benefit. Keep a direct API integration when it already satisfies the task with less maintenance. A single-tool server can still serve a useful client requirement. Tool count alone does not decide the architecture.

Reference: modelcontextprotocol.io

Why use MCP instead of HTTP?

MCP and HTTP are different layers: Streamable HTTP carries MCP messages over HTTP. Choose the MCP message contract when your client needs its defined discovery or invocation behavior. A direct HTTP endpoint may already provide the interface your application needs. Supporting MCP does not automatically provide authorization or stateful sessions.

Reference: modelcontextprotocol.io

Is MCP like Zapier?

MCP is a protocol; Zapier is an automation platform that also offers an MCP interface. A Zap workflow connects configured triggers and actions, while Zapier MCP exposes configured app actions to an AI client. MCP access still requires selecting tools and connecting the relevant accounts. It does not automatically expose every action or remove setup and permissions. Compare the particular workflow you need with the actions available in your configured server, rather than treating a platform-wide app count as a compatibility guarantee.

Reference: docs.zapier.com

What is the difference between MCP and Copilot?

GitHub Copilot is an AI assistant product; MCP is a protocol that supported Copilot features can use to connect to external tools and data. GitHub documents different MCP configuration routes for its IDE integrations, CLI and cloud agent. Connection options and applicable organization policy depend on the surface. Support for MCP does not mean every server is available automatically, or that every Copilot operation uses MCP. Check the setup for the Copilot feature you actually intend to use.

Reference: docs.github.com

What is the difference between MCP and LLM?

An LLM is a model that processes input and generates output; MCP is a communication protocol implemented by applications and servers. In a tool-enabled application, the model may request an operation, and the application sends the request through its configured client and handles the result. MCP is one way to provide that tool interface. It is not the model itself, and it does not independently decide, grant permission or perform an action without the surrounding implementation.

Reference: modelcontextprotocol.io

Is MCP just a JSON?

MCP uses JSON-RPC 2.0 to carry requests and responses. In the 2026-07-28 revision, each request carries its protocol version and client capabilities. The 2025-11-25 revision uses an initialize handshake. When debugging, check which revision both sides support, then inspect the method, input and transport.

Reference: modelcontextprotocol.io

03

MCP Adoption — Who Actually Uses It

MCP support is documented across several products, but clients and servers have different roles. Apple’s Xcode example exposes tools to an external agent; it is not proof of MCP support throughout iOS or macOS.

Highest-value misconception

MCP is not only for Claude.

The MCP project’s December 2025 announcement listed support across several AI products. A platform name is not a compatibility guarantee. Check the actual client surface, supported features, transport and access requirements for the server you intend to connect.

Is MCP only for Claude?

No. MCP is used beyond Claude. The MCP project’s December 9, 2025 announcement listed client support across products including ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code. That is a dated project statement, not a guarantee that each product supports every server feature. For a real connection, check the exact client, transport, configuration and access requirements instead of relying on a platform list.

Reference: blog.modelcontextprotocol.io

Does ChatGPT use MCP?

Yes: the MCP project’s December 9, 2025 announcement listed ChatGPT among products with MCP client support. That is a dated support statement, not proof that a particular server is connected to your ChatGPT account. To evaluate an integration, check the current connection options for the ChatGPT feature you intend to use and verify the server’s supported tools, transport and access requirements.

Reference: blog.modelcontextprotocol.io

Does Microsoft use MCP?

Yes. GitHub documents MCP support in Copilot’s IDE integrations, CLI and cloud agent. The setup differs by surface: IDE connection options depend on the IDE, while cloud-agent configuration can be repository-specific. Applicable organization policy can also control access. This is useful support for an intended integration, not evidence that every Microsoft product or every installed copy of VS Code can call any server without configuration.

Reference: docs.github.com

Is Apple using MCP?

Yes. Apple documents an MCP server that exposes Xcode capabilities to external agents. Its instructions require enabling access in Xcode’s Intelligence settings, configuring the external agent with xcrun mcpbridge, and opening the project in Xcode. This is specific evidence for an Xcode integration; it does not establish that every iOS or macOS AI feature uses MCP.

Reference: developer.apple.com

Is Zapier MCP free?

Zapier lists MCP as available to all accounts, but included access does not mean unlimited task-free execution. As checked on November 7, 2026, Zapier’s usage documentation says there is no separate MCP billing: each successful tool call consumes two tasks from the account’s shared allowance, while failed calls do not consume tasks. Check your current plan and task balance before estimating a workload; availability of a Free account does not remove those limits.

Reference: docs.zapier.com

04

When MCP Breaks Down

A connection can fail because client capabilities, transport or access requirements do not match. Diagnose the actual integration, and keep a direct interface when the intended client does not need MCP.

Practitioner test

Configure authorization for the actual deployment.

The HTTP authorization framework is optional in the base specification; when implemented, it brings concrete discovery and token-validation requirements. Verify those controls in the server you deploy. Local STDIO credentials follow a different path.

When not to use MCP?

Defer an MCP adapter when the intended client does not need it and an existing integration already does the job. An inference-only application with no external capabilities to expose may also have no use for a tool server. If performance is the concern, compare the actual routes under the same workload rather than assuming a universal MCP delay. The guide to when to skip MCP walks through the tradeoffs in more detail.

Reference: modelcontextprotocol.io

Why are people moving away from MCP?

A team may remove an MCP adapter if client compatibility, access-control work or operational maintenance outweighs its benefit. That is a project decision, not proof of a general abandonment trend. When debugging, check the protocol version, transport and tool call separately.

Reference: modelcontextprotocol.io

Is A2A dead?

No. The A2A project documents Agent2Agent communication and describes MCP as complementary. A2A connects agents collaborating on tasks; MCP connects an application or agent to tools and resources. An application may use both, depending on which interaction it needs to implement. Their different roles do not establish a ranking by downloads, a general adoption gap or which protocol will survive.

Reference: a2a-protocol.org

Will MCP replace API?

MCP can sit alongside an existing API rather than replace it. A server might adapt an API, expose database operations or work with files. There is no requirement that every server wrap REST. Decide which interface each intended client needs and keep ownership of the underlying operation explicit.

Reference: modelcontextprotocol.io

Why are people against Copilot?

A team may decide against GitHub Copilot, an AI coding assistant, when its cost, generated output or data-handling controls do not meet that team’s requirements. Those product-evaluation questions are separate from support for MCP, the protocol used to connect a compatible application to external capabilities. Assess the specific Copilot plan and surface, review its current terms and controls, and test the intended coding task. These are evaluation criteria, not a measurement of how common any objection is.

Reference: docs.github.com

05

MCP Architecture & Security

Transport layers, encryption, auth, and the gap between what MCP promises and what your server actually ships.

Security gap

Check the HTTP exposure and security controls.

For Streamable HTTP, validate incoming origins and review authentication and network binding. The HTTP authorization framework requires HTTPS for authorization-server endpoints, with specific redirect-URI rules. STDIO exchanges process streams and is not an HTTP deployment.

Is MCP built on HTTP?

MCP can run over a network or through a local subprocess. The 2026-07-28 revision defines stdio and Streamable HTTP. With Streamable HTTP, a client sends each message in a POST, and the server replies with JSON or an event stream for that request. The 2025-11-25 revision also defined a GET stream endpoint. Check which revision your client and server support.

Reference: modelcontextprotocol.io

Does MCP use HTTP or HTTPS?

Streamable HTTP is the HTTP-based MCP transport; STDIO does not use HTTP. The 2025-11-25 HTTP authorization framework requires HTTPS for authorization-server endpoints and HTTPS or localhost for redirects. Check the authorization and transport requirements for your deployed revision.

Reference: modelcontextprotocol.io

Does MCP use OAuth?

Yes, MCP specifies an OAuth-based framework for HTTP authorization. Supporting the protocol does not mean a deployed server has implemented that framework. Authorization is optional in the base specification; protected operations still need appropriate access controls. STDIO implementations should obtain credentials from their environment rather than follow the HTTP authorization flow.

Reference: modelcontextprotocol.io

Is an MCP server just an API?

An MCP server exposes a particular protocol interface, not a replacement for every kind of API. For a server exposing tools, tools/list returns tool definitions and their schemas; it is not a manifest of every protocol capability. Servers can also expose resources and prompts. MCP messages use JSON-RPC. The 2026-07-28 revision uses self-contained requests. Earlier revisions could establish a stateful HTTP session.

Reference: modelcontextprotocol.io

06

MCP Ecosystem & What's Next

An MCP adapter can connect an existing workflow to an AI client. Check which client needs the tools, who will maintain the connection and how the team will verify its output.

What has replaced Power Automate?

MCP does not replace a workflow product such as Power Automate. MCP defines a client-server interface; a workflow product supplies its own automation features. An application can expose a workflow through a tool, or use a direct integration without MCP. Decide what must happen, which client needs to initiate it, and what permissions and recovery the implementation provides. A protocol comparison alone does not show that an AI agent will handle edge cases better than a configured workflow.

Reference: learn.microsoft.com

Who maintains an MCP scraper connection after setup?

Assign an owner for the MCP client, the scraper server and the source check. The host application runs the client that connects to the server; the server exposes the collection tools. For an agency workflow, ask who will check that the client still lists the required tool, that access works, and that one public source returns the fields needed for the brief. After a client or server update, repeat that small check and record missing fields before using the output. Keep the source URL, capture date and returned coverage with the result. Use the scraper connection walkthrough to plan the first test in your intended client.

Reference: modelcontextprotocol.io

Is MCP basically an API?

MCP is a particular software interface contract, and an MCP server may use an existing API underneath it. What MCP adds is defined capability discovery and invocation for compatible clients. A new tool can be exposed through the server’s tool list, but a client still needs to discover it, support it and have permission to call it. Adding a tool does not guarantee every connected agent notices or uses it automatically. For an existing API, add an MCP adapter when an intended client needs that contract.

Reference: modelcontextprotocol.io

Try one MCP connection.

Connect a compatible client, confirm that it can see the tools, and test one public page before expanding the workflow.

Open the MCP setup walkthrough

MS
Written by

MCP Scraper

Editorial team

About MCP Scraper

See our editorial note for how we prepare guides and handle corrections.

Quick citation guide

Select a citation to automatically copy it to your clipboard.

About our guides · Suggest a correction