MCP specification
MCP specification
Only really needed when building your own MCP server or client.
The full technical specification of the protocol: tools, resources, prompts and how clients and servers communicate.
Overview
The MCP specification is the original, normative document defining what clients and servers must do to be compatible with the Model Context Protocol. It is built on a TypeScript schema, and the current version carries the date 2026-07-28.
The protocol uses JSON-RPC 2.0 messages between three roles: host (the AI application that initiates connections), client (the connector inside the host) and server (the service providing context and capabilities). The idea is inspired by the Language Server Protocol from the developer tools world.
Beyond the core, the specification has optional extensions such as Tasks (running long tasks asynchronously) and MCP Apps (interactive interfaces displayed right in the conversation), along with a dedicated section on security and safety.
Highlights
Official authoritative source
Defines mandatory requirements, using the keywords MUST and SHOULD per BCP 14.
Independent, stateless requests
Each request carries the protocol version and capabilities it needs, so the server does not have to remember earlier requests.
Server-side features
Resources (context data), Prompts (message templates, workflows) and Tools (functions for the model to execute).
Client-side features
Elicitation lets a server ask the user for more information. Sampling has been marked deprecated since version 2026-07-28.
Optional extensions
Tasks, Skills over MCP, MCP Apps; enable them only when both client and server support them.
Clear security principles
Emphasizes user consent, data privacy and safety when calling tools.
What it's for
Building compatible AI applications
A product team uses the specification as the standard when writing a client for their app.
Writing a server without an SDK
When writing in a language with no SDK yet, developers follow the specification to implement correctly.
Evaluating vendors
An IT team uses the specification to check whether an external MCP solution meets security requirements.
Building internal policy
Draw on the consent and privacy principles to write rules for AI use in the business.
Technical training
Serve as the standard reference when training IT staff on how AI connects to systems.
How to use it
- 1
Read the overview
Grasp the three roles host, client, server and the JSON-RPC 2.0 format.
- 2
See the architecture and base protocol sections
Learn how capability discovery works, statelessness, and utilities such as cancellation and progress reporting.
- 3
Read the server and client features
Go deeper into tools, resources, prompts and elicitation depending on whether you are building a server or a client.
- 4
Review the security section
Compare your system against the principles of user consent, privacy and tool safety.
- 5
Check the extensions
Apply Tasks or MCP Apps only when truly needed and both sides support them.
Video guides
The Future of MCP
David Soria Parra, Anthropic
MCP Gets OAuth: Understanding the New Authorization Specification
Microsoft Developer
MCP 2 0 Stateless, New Auth CIMD, and Exciting Extensions by James Ward
Devoxx
Real-world example
A sales management software company in Hanoi wants its customers to be able to use an AI assistant with their order data. The technical lead reads the specification to design the MCP server: orders are exposed as a resource, creating a dispatch slip is a tool, and every tool call must be confirmed by the user, following the safety principles in the specification.
Tips & notes
Always state the version
The specification is named by date; record the version you support in your product documentation.
Avoid deprecated features
For new projects, do not build on Sampling since it is marked deprecated.
Prefer the official SDKs
The SDKs already handle most protocol details; read the specification in depth only when needed.
Treat tool descriptions as untrusted
The specification advises not trusting tool annotations unless the server is trusted.
FAQ
Which version is the current specification?
The specification page points to the schema of version 2026-07-28.
What transport does MCP use?
The data layer uses JSON-RPC 2.0; the transport layer supports stdio and Streamable HTTP.
Does MCP guarantee security by itself?
No. The specification says the protocol cannot enforce security principles itself; implementers must build their own consent and authorization mechanisms.
Do non-technical people need to read it?
Not required, but the security section is very useful for building safe AI usage rules.
Useful links
Want to apply this in your business?
2dm looks at your workflows, picks the right tools and sets them up with your team. The first call is free.
Same topic
Model Context Protocol
4.5
The official MCP documentation site, an open standard for connecting AI assistants to external data and tools.
MCP Servers
4.5
A repository of reference MCP servers and a list of community-built servers.
MCP Registry
3.5
The official directory for finding and publishing MCP servers.