# xmcp v1 is here (/blog/xmcp-v1)

Published: 2026-08-25

xmcp v1 splits the compiler out of the runtime, moves to MCP revision 2026-07-28 through SDK v2, and keeps every existing tool, prompt and resource working unchanged.

What started as a shortcut for building MCP servers in TypeScript is now a complete, production-ready framework with a far smaller footprint.

A major version bump, not a migration.

```bash
npm install xmcp@latest
npm install --save-dev @xmcp-dev/compiler@latest
```

## The runtime stopped carrying the compiler

Through 0.7.1, `xmcp` shipped two kinds of software in one package: the runtime a built server needs to receive requests and run tools, and the compiler that discovers files, invokes Rspack and TypeScript, and produces the build.

Both were necessary. They were not necessary at the same time. Once `xmcp build` finished, the compiler had no work left to do — and production installed it anyway.

v1 splits them. `xmcp` is the runtime. `@xmcp-dev/compiler` is a development dependency.

| Measure                  | 0.7.1      | 1.1.0    | Reduction |
| ------------------------ | ---------- | -------- | --------- |
| Fresh production install | 101.13 MiB | 9.30 MiB | 90.8%     |
| Production dependencies  | 169        | 2        | 98.8%     |
| Runtime tarball          | 4.13 MB    | 1.51 MB  | 63.3%     |

The generated server was already self-contained, and still is. Delete `node_modules` after a build and the HTTP, STDIO, CommonJS, ESM and MCP App artifacts all keep running.

## Current protocol, both client eras

v1 moves to MCP revision [`2026-07-28`](https://modelcontextprotocol.io/specification/2026-07-28) through SDK v2, where `@modelcontextprotocol/server` and `@modelcontextprotocol/client` replace the v1 monolith.

Every transport — HTTP, STDIO, Cloudflare Workers, and the Next.js, Express, Fastify and NestJS adapters — serves both generations. Requests using the 2026 envelope (`server/discover`, `Mcp-Method` / `Mcp-Name` routing, cacheable list results) are handled natively. 2025-era clients go through the SDK's stateless fallback. HTTP stays strictly stateless either way: a fresh server per request, no session cache, no `Mcp-Session-Id`.

Multi Round-Trip Requests replace held-open elicitation. A tool returns `inputRequired(...)` to ask for more before producing a result, and the same tool still serves 2025-era clients — the legacy shim converts it into real elicitation.

Authoring has not changed. A file is still the registration:

```typescript title="src/tools/greet.ts"
import { z } from "zod";
import type { InferSchema } from "xmcp";

export const schema = {
  name: z.string().describe("The name of the user to greet"),
};

export const metadata = {
  name: "greet",
  description: "Greet the user",
};

export default function greet({ name }: InferSchema<typeof schema>) {
  return `Hello, ${name}!`;
}
```

Schemas now register through SDK v2's Standard Schema interface, which keeps the existing Zod 3-or-4 peer range working.

## Sampling in 1.1

`extra.sample()` lets a handler request an LLM completion from the connected client mid-tool, mirroring `extra.elicit()`. It supports text, image and audio messages, `systemPrompt`, `maxTokens`, `modelPreferences`, `temperature` and `stopSequences`.

## Upgrading

Keep the runtime and compiler on matching versions. Your existing tools, prompts, resources, configuration and package scripts work unchanged. v1 requires Node 22 or newer.

```bash
npx create-xmcp-app@latest
```

The [installation docs](https://xmcp.dev/docs/getting-started/installation) cover pnpm, Yarn and Bun. The benchmark fixtures behind the numbers above live in [`packages/xmcp/bench`](https://github.com/basementstudio/xmcp/tree/main/packages/xmcp/bench) — run `pnpm bench` to reproduce them.