The Model Context Protocol (MCP) is an open standard that defines how AI applications connect to external tools and data sources. A system exposes its functions and data once as an MCP server, and any MCP-compatible agent or assistant can discover and use them, instead of each AI product needing its own custom integration.
Anthropic open-sourced MCP in November 2024 to replace fragmented, one-off integrations with a single protocol.1 In December 2025 it donated the protocol to the Agentic AI Foundation, a fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI, and by then ChatGPT, Gemini, Microsoft Copilot, and Visual Studio Code had adopted it.8
For a company, MCP changes the integration question. Instead of wiring an order system separately into every AI tool, you build or install one MCP server for it, and decide in one place what any AI agent may read or do there.
Hosts, clients, and servers
MCP has three participants.3 The host is the AI application, such as a support agent or a desktop assistant. For every system it connects to, the host creates a client, a connector that keeps one connection to one server. The server is a program that exposes a system's data and functions, whether it runs on the same machine or remotely.
The design borrows from the Language Server Protocol, which let any code editor support any programming language through one standard interface.2 MCP does the same for AI applications and the systems they need.
Tools
Functions the model can call to act: query a database, check a delivery, create a ticket. Tools are model-controlled, meaning the model decides when to use them.4
Resources
Data the application can read for context: file contents, database records, a schema, a policy document.3
Prompts
Reusable templates a server offers for common tasks, such as a structured way to summarize a claim or draft a reply.3
Elicitation
A way for a server to ask the user for missing input or confirmation, for example before it changes a booking.2
A real tool call, message by message
A customer asks a retailer's support agent where an order is. The model decides to use the order server's status tool. The host's MCP client sends this JSON-RPC 2.0 request, in the shape the current specification defines.47
{
"jsonrpc": "2.0",
"id": 7,
"method": "tools/call",
"params": {
"name": "get_order_status",
"arguments": { "order_id": "48213" },
"_meta": {
"io.modelcontextprotocol/protocolVersion": "2026-07-28",
"io.modelcontextprotocol/clientInfo": { "name": "support-agent", "version": "1.4.0" },
"io.modelcontextprotocol/clientCapabilities": {}
}
}
}The server checks the caller's authorization, runs the lookup against the order system, and returns the result as content the model can read.
{
"jsonrpc": "2.0",
"id": 7,
"result": {
"resultType": "complete",
"content": [
{ "type": "text", "text": "Order 48213: shipped 11 Nov, last scan 13 Nov at the regional depot, status delayed." }
],
"isError": false
}
}Illustrative values (tool name, order number, client name) in the message format of the 2026-07-28 specification. The model then writes the reply to the customer from this result. Illustrative example. The tool name, order number, and dates are invented; the message shape follows the specification.
MCP, function calling, APIs, and A2A
MCP sits between a model's ability to call functions and the systems' own APIs. It does not replace either.
| Term | What it standardizes | Between | In a support stack |
|---|---|---|---|
| Function calling | How a model asks its application to run a function | A model and its own application | The model emits get_order_status with arguments |
| Model Context Protocol | How applications discover and call tools and data in other systems | An AI application and many systems | The order system's MCP server, usable by any compatible agent |
| REST or GraphQL API | How software reads and writes one system | Any two programs | The order system's API that the MCP server calls underneath |
| Agent2Agent (A2A) | How agents delegate tasks to other agents | Two agents | A support agent handing a claim to a separate claims agent |
Where companies use MCP
MCP pays off where several agents need the same systems, or where one agent needs many of them.
One agent, many back-end systems
A telecom's support agent reaches billing, plan changes, network status, and the helpdesk through four MCP servers instead of four custom integrations.
list tools → call billing → change plan → logCatalog and stock as tools
An electronics retailer exposes product search, live stock, and delivery dates as tools for its shopping agent and its in-store staff assistant.
search → check stock → add to cartControlled access to the ERP
An insurer's claims agent reads policies as resources and may only create draft entries in the ERP, never approve payments.
read policy → match claim → draft entryOne gateway to company documents
A knowledge agent searches contracts, procedures, and past tickets through servers that respect each employee's existing access rights.
search → filter by rights → citeIllustrative examples, not client deployments.
Security considerations and controls
The specification says plainly that MCP enables arbitrary data access and code execution paths, and that the protocol itself cannot enforce its security principles. Implementers must.2
Untrusted tool descriptions
ControlA malicious or compromised server can describe its tools misleadingly. Clients must treat tool annotations as untrusted unless the server is trusted.4 Install only vetted servers and pin their versions.
Prompt injection through results
ControlTool results and resources can carry text that tries to redirect the model.9 Keep humans able to deny tool calls, as the specification advises, and confirm high-impact actions.4
Token passthrough
ControlA server that forwards a client's token to a downstream API without checking it was issued for that server opens unauthorized access. The security guide calls this an anti-pattern.6
Over-broad scopes
ControlRequest the minimum scopes per server and per tool, and let the server return only the tools the caller's authorization permits.46
Local server compromise
ControlA server installed on a laptop can reach that machine's files and network. Run servers in managed environments and review what each one can access.6
Common questions
Who owns the Model Context Protocol?
Anthropic created it and released it as open source in November 2024. In December 2025 it donated MCP to the Agentic AI Foundation under the Linux Foundation, which now stewards it as a vendor-neutral standard.
Is MCP only for Claude?
No. It is an open protocol with SDKs in the major programming languages. ChatGPT, Gemini, Microsoft Copilot, Cursor, and Visual Studio Code have adopted it, and any application can implement a host, client, or server.
What is an MCP server?
A program that exposes a system's functions and data in MCP format. For example, an order-management MCP server could offer tools to check status and start a return, plus resources such as the return policy.
Does MCP replace APIs?
No. An MCP server usually calls the system's existing API underneath. MCP adds a standard way for AI applications to discover what is available and call it, with authorization handled in one place.
Is MCP secure enough for customer data?
The protocol supports OAuth-based authorization and the specification sets consent and access-control principles, but security depends on the implementation. Scope permissions narrowly, vet every server, and log every tool call.
References
- Anthropic. Introducing the Model Context Protocol. November 2024.
- Model Context Protocol. Specification, version 2026-07-28.
- Model Context Protocol. Architecture overview.
- Model Context Protocol. Specification: Tools. Version 2026-07-28.
- Model Context Protocol. Key changes in 2026-07-28.
- Model Context Protocol. Security best practices.
- JSON-RPC Working Group. JSON-RPC 2.0 Specification. 2010.
- Anthropic. Donating the Model Context Protocol and establishing the Agentic AI Foundation. December 2025.
- OWASP GenAI Security Project. LLM01:2025 Prompt Injection. OWASP Top 10 for LLM Applications, 2025.