Skip to main content

Model ContextProtocol (MCP)

A printer’s case of metal type sorted into compartments, labelled Scotch Roman (Pica) Caps

MCP is an open standard for connecting AI models to external tools and data through a uniform server interface — the transport layer beneath most production agent tool use.

Generative AI. Last reviewed August 2026.

Ask an engineerAll terms

In this entry

Long tables of catalogue cards being sorted in the Card Division of the Library of Congress, early twentieth century

5sections

Definition

The Model Context Protocol (MCP) is an open standard for connecting AI models to external tools, data sources and services through a uniform interface. An MCP server exposes a set of capabilities — tools to call, resources to read, prompts to reuse — and any MCP-aware client can consume them without a bespoke integration.

The practical value is combinatorial. Without a shared protocol, connecting N assistants to M systems is N × M integrations. With MCP it is N + M.

Where it fits in an agent stack

MCP is the transport and interface layer for tool use. It is not an agent framework, not a runtime, and not a model.

On AWS, Amazon Bedrock AgentCore Gateway is the governed entry point through which an agent reaches tools — including MCP servers — with authentication, customer-configurable rate limits, and Cedar-based fine-grained access control on writes. AWS also publishes its own MCP servers for interacting with AWS services.

MCP in commerce

For merchants, an MCP server over the catalog is how an agent reaches product, inventory and order data without you shipping a separate integration for every assistant that appears.

One surface to maintain, one place to enforce authentication and rate limits, one audit log. That framing matters, because an MCP server exposing commercial data is a public interface — an infrastructure and security design problem before it is an AI one. Blast radius under a scrape, per-caller rate limiting, and what a compromised token can reach are the questions to settle first.

When not to use it

  • A single, stable integration. If exactly one assistant needs exactly one system, a direct API call is simpler and has fewer moving parts. MCP earns its keep at N × M, not at 1 × 1.
  • Write-heavy, high-risk actions without an authorization layer. MCP describes what a tool does; it does not decide whether this caller may do it. Pair it with policy evaluation outside the model — Cedar via AgentCore Gateway, or your own authorization service.
  • As a substitute for a data contract. Exposing four systems that disagree about what a product identifier is, through a clean protocol, produces confidently wrong answers with better formatting.

AI agent · Tool catalog · Amazon Bedrock AgentCore · Agentic commerce · Human in the loop

The catalogue room of the Bibliothèque Mazarine in Paris, wooden card cabinets under a painted ceiling
Catalogue room, Bibliothèque Mazarine, Paris

More in Generative AI

Related services

  • Amazon Bedrock AgentCore ConsultingProduction agent infrastructure on Amazon Bedrock AgentCore — Runtime, Gateway, Memory, Identity, Observability and Evaluations. Bring your own framework; we operate the substrate.
  • Agentic Commerce ReadinessMake your store sellable to AI shopping agents. ACP and UCP protocol surfaces, an MCP server over your catalog, and product data that survives an agent comparison — built on AWS.
  • Amazon Bedrock Consulting for Production LLM ApplicationsAmazon Bedrock implementation consulting — Knowledge Bases, Agents, Guardrails, model routing, and production RAG. Hands-on Bedrock engineering, not GenAI strategy.

Reading

Ask

Our AWS-certified team implements, audits and tunes Model Context Protocol (MCP) in production. Tell us what you are building.

Pictures:James Murray in the Scriptorium (public domain); Card Division of the Library of Congress, 1900–1920 (public domain); Salle des catalogues, Bibliothèque Mazarine, Paris, Marie-Lan Nguyen (CC BY 2.0); Copyright card catalog drawer, Michael Holley Swtpc6800 (public domain); Type in sort case, Edinburgh City of Print (CC BY 2.0); Widener Library reading room, 1920 (public domain). Via Wikimedia Commons.