Standardization
A consistent protocol for discovery, capability negotiation, and message exchange.
Open standard · Built for agents
One protocol.
Every connection.
Model Context Protocol gives AI applications a shared way to connect with tools, data, and workflows. Build the capability once and use it everywhere.
Open source · JSON-RPC 2.0 · Local or remote
mcp / live topology
connected
Reasoning layerAI model
Protocol layerMCP client
Adapter AGitHub
Adapter BDatabase
Adapter CFiles
tools
resources
prompts
TOOLSRESOURCESPROMPTSAUTHORIZATIONSAMPLINGCOMPLETION
The integration layer AI was missing
Stop rebuilding the same connection.
Before MCP, every AI application needed a bespoke adapter for every tool and data source. MCP replaces that mesh with a stable, shared contract.
Without MCP
N × M
Every client connects to every integration.
With MCP
M + N
Each side implements the protocol once.
Clear security boundaries
Defined authorization flows and consent expectations help implementations make access explicit.
Interoperability
Move between models, hosts, tools, and infrastructure without rewriting the core interface.
Developer experience
Official SDKs, reusable servers, and inspectable primitives make integrations faster to ship and easier to debug.
A small protocol with clear roles
The model reasons.
The client connects.
Servers provide.
An MCP host manages one or more clients. Each client maintains a focused connection to a server, exchanging JSON-RPC messages over local or remote transports.
Explore the architecture
Model role / Reason
Model + Host
Chooses when context or an action is needed.
Client role / Connect
MCP Client
Negotiates capabilities and routes messages.
Server role / Provide
MCP Servers
Tools
Resources
Prompts
initialized
capabilities negotiated
protocol / 2026-07-28
Three primitives.
A world of capability.
The vocabulary stays compact so integrations can stay composable.
Tools
Executable functions the model can invoke to retrieve information or take action.
tools/call
Resources
Structured context such as files, records, schemas, or state, managed by the application.
resources/read
Prompts
Reusable, server-provided templates that shape repeatable workflows.
prompts/get
An open standard, broadly adopted
Built across the ecosystem.
MCP is supported by model providers, developer platforms, and cloud infrastructure, creating a portable foundation instead of another closed integration layer.
“Build once and integrate everywhere” is no longer an aspiration. It is the operating model of a growing, vendor-neutral ecosystem.Join the community ↗
The practical questions
MCP, without the mystique.
What teams usually want to know before they wire up their first server.
What exactly is MCP?
Model Context Protocol is an open standard for connecting AI applications to external systems. It defines how clients and servers discover capabilities and exchange resources, prompts, and tool calls.
Is MCP a model or an agent framework?
Neither. MCP is the connection layer. It works alongside your model, agent framework, and application architecture rather than replacing them.
Does MCP make integrations secure by default?
MCP defines authorization capabilities and emphasizes user consent and control, but implementers remain responsible for access control, data protection, tool safety, and clear approval experiences.
Can MCP servers run locally and remotely?
Yes. Standard transports cover local process communication over stdio and remote connections over Streamable HTTP.
Do I need to replace my existing APIs?
No. An MCP server can sit in front of existing APIs and systems, exposing selected capabilities through a consistent model-facing interface.
Where should I start?
Start with the official architecture overview, then build a small server around one useful tool or resource. Use the Inspector to verify capabilities before connecting it to a production client.
Open the connection
Your first MCP server is closer than you think.
Start with the protocol, choose an SDK, and expose one capability. The ecosystem takes it from there.