- Type
- Open standard / protocol
- Introduced by
- Anthropic
- Released
- November 2024
- Governed by
- Agentic AI Foundation (Linux Foundation), from December 2025
- Key use
- Connecting AI models to tools and data sources
- Related
- AI agents, Tool use, Function calling, Agentic AI
- Type
- Open standard / protocol
- Introduced by
- Anthropic
- Released
- November 2024
- Governed by
- Agentic AI Foundation (Linux Foundation), from December 2025
- Key use
- Connecting AI models to tools and data sources
- Related
- AI agents, Tool use, Function calling, Agentic AI
The Model Context Protocol (MCP) is an open standard that defines a common interface for large language models (LLMs) and AI agents to communicate with external tools, data repositories, and services. Introduced by Anthropic in November 2024, MCP addresses a fragmentation problem in the AI tooling landscape: before its introduction, each AI application required bespoke integrations with every external system it needed to access, resulting in duplicated effort and brittle pipelines. MCP provides a single, standardised protocol that any AI system and any external service can implement once, enabling interoperability across the ecosystem.
Background and Motivation
As AI systems evolved from stateless question-answering interfaces to agentic applications capable of taking sequences of actions, the need to connect models to real-world systems became acute. Developers building AI assistants needed to integrate with databases, APIs, file systems, communication platforms, calendar services, and countless other tools. Without a standard, every integration was a point-to-point custom implementation, requiring the AI application developer to understand the specific API of every service and the service developer to know nothing about the AI application.
MCP draws an analogy to the USB standard in hardware: just as USB allowed any device to connect to any host without custom cables, MCP allows any AI application to connect to any tool server without custom glue code.
Technical Architecture
MCP follows a client-server architecture with three primary roles.
The host is the application in which the AI model runs — an IDE, a chat interface, or a custom AI workflow. The client is the MCP client library embedded in the host, which manages connections to MCP servers. The server is a lightweight service that exposes a specific capability — for example, a file system server, a web search server, or a database server.
Servers expose capabilities through three primitive types. Tools are functions the model can invoke to perform actions, such as executing a database query or sending an API request. Resources are data objects the model can read, such as file contents or structured records. Prompts are server-provided instruction templates that shape how the model interacts with the server's domain.
The protocol uses JSON-RPC 2.0 as its message format and supports two transport mechanisms: standard input/output for local processes and HTTP with Server-Sent Events for networked services. SDKs were released for Python, TypeScript, C#, and Java.
Sampling and Agentic Loops
MCP also defines a sampling primitive that allows servers to request completions from the host model, enabling multi-hop agentic workflows in which servers can themselves trigger model reasoning. This bidirectional capability distinguishes MCP from simpler tool-calling conventions and enables complex orchestration patterns.
Industry Adoption
Following its release, MCP gained rapid adoption across the AI industry. In March 2025, OpenAI adopted MCP and integrated it into ChatGPT desktop and API products. Google DeepMind confirmed MCP support in Gemini models in April 2025. Microsoft, Cloudflare, Block, and numerous independent tool providers released MCP servers for their services.
By mid-2025, the Python and JavaScript SDKs for MCP servers were recording over 20 million weekly downloads. The ecosystem grew to include thousands of community-contributed MCP servers covering tools ranging from GitHub and Slack to PostgreSQL, Stripe, and proprietary enterprise APIs.
In December 2025, Anthropic donated MCP governance to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI. This move placed MCP on a neutral governance footing, reducing concerns about vendor lock-in and accelerating enterprise adoption.
Comparison with Alternative Approaches
Prior to MCP, the dominant approach to connecting AI models to tools was function calling — a pattern supported by OpenAI, Anthropic, and Google APIs where developers define tool schemas in JSON that the model can invoke during generation. Function calling remains the mechanism by which a model decides to use a tool; MCP standardises the layer between the AI application and the tool implementations themselves. The two are complementary rather than competing.
OpenAPI specifications and plugin systems (including early ChatGPT Plugins) addressed related problems but lacked the bidirectional agentic primitives and the formal governance structure that MCP provides.
Malaysian enterprises adopting AI-powered workflows stand to benefit directly from MCP. The protocol reduces the engineering effort required to connect AI models to internal systems — a significant consideration for organisations where AI development capacity is limited. Banks such as Maybank and CIMB, and telecommunications providers such as TM and Maxis, operate large technology stacks including core banking systems, CRM platforms, and data warehouses that AI agents need to query and act upon.
MDEC and the National AI Office Malaysia have both noted that reducing integration friction is a prerequisite for broader AI adoption in the Malaysian enterprise market. MCP-compatible tool servers for common Malaysian enterprise platforms — including SAP and Oracle ERP systems widely used by Petronas, AirAsia, and large public-sector organisations — lower the barrier to deploying AI agents in these environments.
The AWS, Azure, and Google Cloud infrastructure available in Malaysia, all of which have expanded their MCP-compatible managed AI services, means that Malaysian organisations can adopt MCP-based architectures without building protocol infrastructure from scratch. Malaysian AI companies and system integrators, including those in the Cyberjaya and TechCity KL ecosystems, are building MCP servers tailored to local regulatory requirements and Malaysian-specific data sources such as EPF, LHDN, and BNM APIs.
HRD Corp-funded training programmes in Malaysia increasingly include AI agent development curricula. MCP has become a standard topic in these programmes because of its role in production-grade agentic systems, equipping Malaysian developers with transferable skills applicable across the growing ecosystem of MCP-compatible AI platforms.
See Also