Independent Model Context Protocol research

Model Context
Protocol

How MCP connects AI applications to tools and data, with guides to building servers, choosing transports, and handling permissions.

Protocol statusCurrent
202607.28

Requests carry their own context. Servers advertise operations through discovery, and Tasks is available as an extension.

Read the annotated change guide

Architecture decisions

Hosts, servers and capabilities

A host runs the AI application. Its MCP client connects to servers that provide tools, data and prompts.

02 / Where should it run?

Local and remote servers

Local servers run on your machine; remote servers are reached over a network. Compare file access, credentials and maintenance.

Choose local or remote MCP
03 / What should it expose?

Tools, resources and prompts

Tools perform operations, resources supply data, and prompts prepare messages. Examples show when to combine them.

Compare tools, resources, and prompts

For existing servers, check protocol versions or review MCP security. For an introduction, read the plain-language MCP overview.

01

July 2026 revision

MCP 2026-07-28 changed how clients discover and call servers.

Before July 2026, an MCP connection began with initialization and capability negotiation. The current core is stateless: requests carry their own context, servers advertise operations through discovery, and formerly core features can live as extensions.

Older clients may still use initialization and sessions. The version comparison explains which message flows changed and what a migration needs to account for.

MCP guides

Definitions, code examples, protocol changes and security controls.

Tested Python guideSpec 2026-07-28

Build an MCP Server in Python

Build a local MCP server with Python, expose one typed tool, verify its schema and errors, and call it through an actual stdio subprocess.

Read guide

History

From first public commit to stateless core

Explore the full timeline
First public repositories

Specification and SDK history begin before the public launch.

MCP launches

A stateful, JSON-RPC-based protocol with tools, resources, prompts, and more.

Remote transport matures

Streamable HTTP and a protocol authorization framework arrive.

Neutral foundation

MCP joins the Agentic AI Foundation under the Linux Foundation.

Stateless core

Stateless core, discovery, MRTR, extensions, and explicit deprecations.

References

MCP source library

Search 551 records covering specifications, SDKs, repository history, security advisories and research.