Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open standard that defines how AI models communicate with external tools, data sources, and services. Think of it as the universal adapter layer between an AI agent and the rest of your digital environment — a single, well-defined interface that replaces a patchwork of bespoke integrations.

Before MCP, connecting an AI model to a new tool meant writing custom code for every combination of model and integration. With ten models and twenty tools, that is two hundred integration points to build and maintain — what engineers call the M×N problem. MCP collapses that to M+N: each model implements the protocol once, and each tool exposes one MCP server. Any model can then reach any tool.

MCP is sometimes described as "the API for AI." Where REST APIs define how software talks to software, MCP defines how intelligent agents talk to the world.

Anthropic introduced MCP in November 2024 as an open-source specification. The decision to open-source the standard from day one was deliberate: a proprietary protocol would fragment the ecosystem; a shared standard would lift all boats.

The protocol drew immediate attention from across the AI industry. Within months, Microsoft, Google, OpenAI, and a growing roster of infrastructure and tooling vendors had either adopted MCP or announced support for it. Developer communities on GitHub published hundreds of community-built MCP servers covering everything from databases and file systems to CRMs, communication platforms, and developer toolchains.

In early 2025 the MCP specification was transferred to a vendor-neutral governance body, cementing its role as a genuine industry standard rather than a single-vendor solution. Today it is the de-facto protocol for agent-to-tool communication across the AI ecosystem.

MCP uses a client/server model. An MCP host (typically an AI application or agent runtime) connects to one or more MCP servers, each of which exposes capabilities to the model. The protocol defines three core primitives:

Tools

Executable functions the model can invoke — for example, searching a database, calling an API, sending a message, or running a calculation. Tools are the action layer of MCP.

Resources

Structured data the model can read for context — documents, records, files, configuration, or any information that grounds the model in the specifics of a situation. Resources are the knowledge layer.

Prompts

Pre-defined, reusable interaction templates that guide how a model approaches a particular task. Prompts allow organisations to encode domain expertise and safe operating patterns directly into the protocol layer.

Communication between client and server can run over stdio (for local processes), server-sent events (SSE), or the newer Streamable HTTP transport — a more robust mechanism suited to cloud-hosted, enterprise-grade deployments. Authentication is handled via OAuth 2.0, making MCP compatible with existing enterprise identity infrastructure.

Enterprise Data Access

Expose internal databases, knowledge bases, and document repositories to AI agents without custom integration code. An agent can query a policy document, retrieve a customer record, and check a project status — all through a single, authenticated protocol.

Multi-Tool Agent Workflows

Complex tasks often require an agent to chain actions across multiple systems: look up a record, draft a response, log a note, and trigger a workflow. MCP gives the agent a consistent interface to all of these, regardless of the underlying technology stack.

Real-Time Context Injection

Rather than relying solely on training data, agents can pull live context — current prices, recent news, updated policies, or user-specific information — at inference time. This keeps responses accurate and relevant without retraining the model.

Safe, Governed AI Integration

Because MCP servers act as a controlled gateway, organisations can enforce access policies, rate limits, and audit logging at the protocol layer — giving them visibility and control over what their AI systems can see and do.

Organisations that build AI capabilities on proprietary integration approaches face compounding technical debt: every new model or tool requires custom work, and the sprawl becomes unmanageable quickly. MCP offers a way out of that trap.

Standardising on MCP at the infrastructure layer yields several strategic benefits. It reduces integration costs and accelerates time-to-value for new AI use cases. It future-proofs the investment — when a better model arrives, swapping it in does not require rebuilding integrations. It enables vendor diversity, avoiding lock-in to a single AI provider. And it creates a reusable asset: every MCP server built for one use case is immediately available to all other agents on the platform.

MCP is also the foundation for agentic architectures. As organisations move beyond single-turn chat towards autonomous agents that plan and execute multi-step tasks, the ability for those agents to reliably discover and invoke capabilities becomes essential. MCP provides exactly that foundation.

For organisations beginning their AI journey, implementing MCP early — before the integration sprawl sets in — is one of the highest-leverage architectural decisions they can make.

The rise of AI agents as active participants in enterprise systems represents a fundamental shift in who (and what) organisations design for. At Exco Partners we describe this shift as Machine Experience — or MX.

MX is the practice of designing systems, data structures, and interfaces not only for human users but for the machines that increasingly navigate, evaluate, and act within those systems. Just as UX demanded that interfaces be clear, empathetic, and usable for people, MX demands that systems be structured, labelled, and discoverable for agents.

MCP is the infrastructure layer that makes MX possible. When an AI agent needs to access a resource, invoke a capability, or retrieve context, it does so through MCP servers. The quality of those servers — how well they describe their capabilities, how reliably they return structured data, how consistently they handle authentication — directly determines the quality of the machine experience.

Designing good MCP servers is, in essence, designing for machine users. That means thinking about the semantics of tool descriptions (can the agent understand what this tool does and when to use it?), the structure of returned data (is it machine-readable and well-labelled?), and the reliability of the integration (does the machine get a consistent, predictable experience?).

Organisations that invest in MX — and in the MCP infrastructure that underpins it — are positioning themselves to participate fully in the agentic economy. Those that do not risk becoming invisible to the AI systems their customers and partners increasingly rely on.