The Model Context Protocol (MCP) is an open-standard communication protocol that allows artificial intelligence models to securely connect with external data sources, tools, and environments. By establishing a standard JSON-RPC 2.0 format, the MCP protocol acts as a universal connector, similar to a USB-C cable for AI, that links a large language model with local files, remote databases, and cloud services. It replaces the complex process of writing custom, fragile API integrations for every single AI application.
According to a 2026 software industry survey by WorkOS and Stacklok, 41% of organisations have successfully deployed MCP servers in production environments, and 72% plan to increase their usage year-over-year. Developer adoption has grown rapidly, with TypeScript and Python SDKs for the protocol maintaining roughly 97 million monthly downloads by early 2026.
What is MCP in AI?
The Model Context Protocol (MCP) is an open-standard, model-facing protocol built on JSON-RPC 2.0 that standardises how generative AI systems securely interface with local or remote databases, files, and tools. Understanding the MCP meaning in AI requires looking at the history of AI integration. Traditionally, developers had to build a bespoke API connector every time they wanted an LLM to read a new database or access a company repository.
To solve this fragmentation, the creators behind MCP, Anthropic, released the first specification in November 2024. In late 2025, Anthropic donated the Model Context Protocol to the newly established Agentic AI Foundation under the Linux Foundation. This open-source approach, backed by technology giants including OpenAI, Microsoft, Google, AWS, and Bloomberg, has established MCP as a foundational standard for AI agents.
The role of MCP clients
In Model Context Protocol architecture, the host application (such as VS Code or Cursor) acts as the primary interface that manages the user session and routes requests. The MCP client resides within the host to establish and handle direct connections to individual servers.
The client translates the host's instructions into the standard protocol format. Many integrated developer environments establish these secure sessions using standard input/output (STDIO) to communicate locally with tools.
The role of MCP servers
An MCP server is a lightweight application that exposes specific resources, templates, and tools to the client. Instead of writing custom APIs, developers build a single server that adheres to the protocol specification.
These servers can run locally or in secure remote cloud environments. This design simplifies deployments, making it easier to scale AI capabilities across a business.
Data sources and external resources
The final piece of the puzzle is the actual data layer. The server translates requests from the client into specific queries for underlying data sources, such as Postgres databases, Slack workspaces, or local file systems. This keeps the AI model completely isolated from direct access to the raw data, ensuring a secure and controlled environment.
Key primitives of the MCP architecture
The entire protocol is designed around three fundamental building blocks, or primitives. These primitives give the AI model everything it needs to safely read data, format instructions, and perform actions.
Resources as data context
Resources are read-only data sources that the server makes available to the AI model. This can include raw text files, database schemas, active error logs, or application state data.
A common mistake is assuming that implementing the protocol eliminates the need for Retrieval-Augmented Generation (RAG). However, the protocol is a communication format, not a database. Developers typically wrap RAG functions within an MCP server to standardise how the model queries vector repositories.
Prompts as template patterns
Prompts are pre-defined templates that help users and models interact effectively. They allow servers to supply standard instructions, shortcuts, and contextual frameworks directly to the client. For example, a server designed for software debugging might supply a prompt template that automatically formats error logs and code snippets before sending them to the host application.
Tools for executing actions
Tools allow AI models to perform actual work. Unlike read-only resources, tools are executable actions that can modify state or trigger external workflows. For example, a tool might allow an agent to write code to a local file, create a GitHub pull request, or send a Slack message.
Security specialists at firms like ThoughtWorks warn that this capability introduces new vectors of risk, specifically regarding prompt injection and tool shadowing. This security vulnerability occurs when a malicious actor registers a rogue server under a name identical to a trusted internal tool, tricking the LLM into executing harmful actions.
Why Model Context Protocol matters for AI agents
Before MCP, building an autonomous agent felt like reinventing the wheel. Every developer had to write custom code to connect their chosen model to a database, and that code became useless if they switched to a different model. The protocol completely solves this fragmentation by offering a single, open standard.
For developers, this means write-once, run-anywhere compatibility. Once you write an MCP-compliant server for a database, any compliant client can connect to it immediately. This allows businesses to build modular AI ecosystems where databases, tools, and models can be swapped out without breaking the system.
Furthermore, standardisation improves security. Because all communication happens over a strict, structured protocol, security teams can easily monitor, audit, and restrict what tools and files an AI model can access. This is particularly crucial in highly regulated environments.
Common use cases and implementations of MCP
Organisations around the world are adopting the protocol to build safer and more capable AI workflows. As highlighted in a Bank of England staff blog post, platforms in the UK fintech and banking sectors use MCP servers to securely link AI-driven open banking agents to transactional databases. This setup lets an autonomous agent securely compile KYC records or trigger payments while fully adhering to the UK Payment Services Regulations.
An important note on terminology: While Model Context Protocol dominates technical discussions in AI, the abbreviation MCP also refers to other concepts across different industries.
- Multi-currency pricing (MCP): In cross-border e-commerce, this is a payment solution that lets international shoppers view prices and complete checkouts in their local currency while the merchant receives settlement and final payouts in their own domestic base currency, such as GBP. Payment processor Nuvei notes that this differs from Dynamic Currency Conversion (DCC), which converts currency at checkout with unpredictable, acquirer-controlled markups. Research from ConnectPay shows that 92% of international consumers prefer to shop on online stores that display localized pricing, and 49% will abandon their shopping carts if local currency options are not available. Implementing this solution can provide e-commerce merchants with an immediate transaction conversion rate boost of 28% to 30%, according to Swell. Outbound cross-border e-commerce from the UK is heavily dominated by fashion and apparel, representing 44% of international online purchases from UK stores, followed by beauty products at 17%, according to the International Trade Administration.
- Mobile contactless payments (MCP): In European retail, this term describes face-to-face transactions via smartphones or wearables. While European payment frameworks use this term, the UK Payment Systems Regulator (PSR) primarily refers to this technology as Contactless Mobile Payments (CMPs).
- Medium combustion plant (MCPD): Under UK environmental policy managed by DEFRA and the Environment Agency, operators of combustion units with a net thermal input between 1 and 50 MW must register, apply for permits, and stay within strict air emission limit values.
For software developers and enterprise architects, however, MCP remains the definitive standard for connecting LLMs to the real world.