MCP 2026-07-28対応のMCPサーバーをTypeScript SDK v2で作ってみた（実装編）This article walks through building a working MCP server compliant with the…

The Model Context Protocol (MCP) continues to move quickly, and a new post on Qiita documents the hands-on process of building a working MCP server that conforms to the MCP 2026-07-28 specification using the TypeScript SDK v2. For developers connecting large language models to external tools and data, tracking the latest spec revision alongside the matching SDK release matters, because interoperability between a client and a server depends on both sides implementing the same protocol version.

The article is explicitly framed as an implementation installment (実装編) that follows an earlier post summarizing the changes introduced in the 2026-07-28 revision. Where the first piece organized what changed at the specification level, this second part turns to practice, walking through the construction of a single runnable server with concrete code examples. That structure — a spec overview followed by a build-along guide — is a common and practical way to absorb a protocol update, since reading a specification and actually wiring up a server exercise different kinds of understanding.

For readers unfamiliar with the ecosystem, MCP is an open protocol, originally introduced by Anthropic in late 2024, that standardizes how AI applications supply and consume context. Instead of every application building bespoke, one-off integrations, MCP defines a client-server architecture in which a host application — such as an AI assistant or an editor extension — communicates with one or more MCP servers that expose capabilities. The primitives a server commonly provides are tools (functions the model can invoke), resources (data the model can read), and prompts (reusable message templates). Under the hood, messages are exchanged using JSON-RPC.

The TypeScript SDK is one of several official language SDKs, alongside implementations for Python and other languages, that handle the protocol's message plumbing so developers do not have to manage the low-level exchange by hand. A jump to a v2 major version typically signals a restructured API surface or breaking changes relative to earlier releases, which is one reason a guide written specifically against v2 is useful: code samples and setup steps from older tutorials may no longer apply cleanly. Readers migrating an existing server would likely need to review the SDK's changes and the spec's changes together.

MCP specifications are identified by date-based labels, and the 2026-07-28 string denotes a particular revision of the standard. This versioning approach lets clients and servers negotiate a shared protocol version during the initial handshake, and it makes the compatibility target explicit. Building against a named revision, as the article does, is a way to ensure that the server advertises and honors the behavior expected by clients targeting that same version.

Although the post's specific code is not summarized in detail, a typical MCP server build in TypeScript involves initializing the server object, declaring its capabilities, registering handlers for tools or resources, and connecting it to a transport. Servers generally communicate over stdio when run as a local subprocess, or over a streamable HTTP transport for remote or networked deployments. The choice of transport, along with correct capability declaration, tends to be where implementation details matter most for real-world compatibility.

The broader context is an increasingly crowded field of MCP tooling. Since the protocol's release, a range of clients — including AI assistants, code editors, and desktop applications — have added MCP support, and public registries and directories of community servers have grown. That momentum makes conformance to the current spec and SDK meaningful, because a server that adheres to the published revision is more likely to work across the expanding set