mcp-apps-builder
Hướng dẫn tham khảo bắt buộc để xây dựng máy chủ MCP sản xuất với các công cụ, tài nguyên, lời nhắc và widget. Bao gồm các khái niệm nền tảng (công cụ, tài nguyên, lời nhắc, nguyên thủy widget), kiến trúc máy chủ, các mẫu xác thực (OAuth, Supabase, tùy chỉnh) và chiến lược triển khai. Có hướng dẫn chi tiết để triển khai công cụ, tài nguyên, lời nhắc, định dạng phản hồi và giao diện người dùng dựa trên widget với quản lý trạng thái và chủ đề. Ghi lại các phản mẫu phổ biến (thiếu xác thực, lỗi không đúng cách...
npx skills add https://github.com/mcp-use/mcp-use --skill mcp-apps-builderBuild MCP Apps with mcp-use v2
Treat the installed mcp-use package, its generated types, and the project's existing exports as the source of truth. Check the installed version before changing code; do not assume APIs from mcp-use v1.
Workflow
- Inspect
package.json, the server entry, exported tool refs,views/, and the installedmcp-useversion. - Scaffold a new project with
npx create-mcp-use-app@latest; do not hand-build framework boilerplate. - Read only the references needed for the task:
- Server primitives for tools, resources, prompts, middleware, and result envelopes.
- Views for interactive MCP Apps and React hooks.
- Authentication for OAuth providers and authenticated tool handlers.
- Migration when converting v1 code or reviewing package boundaries.
- Verification before reporting completion.
- Implement against the package types. Export every statically declared tool ref that a View calls.
- Validate through the real lifecycle: build/typecheck, run the server, connect a client, call the tool, and render the View when one exists.
Native v2 invariants
- Import server APIs from
mcp-use; provider adapters come frommcp-use/oauth/*; React APIs come frommcp-use/react. - Define tools with
inputSchema; addoutputSchemawhen returning structured data or binding a View. - Return MCP result envelopes with
content,structuredContent, and optionally_metaorisError. - Put each View at
views/<name>/view.tsxand bind it withview: { name: "<name>" }. - Read the rendering call with
useToolContext; use focused hooks such asuseCallTool,useViewState,useHostContext, anduseDisplayModefor additional behavior. - Export the server as the default export. Let
mcp-use dev,build, andstartown framework lifecycle and View compilation. - Keep request state in the request context or an external store. Do not rely on module globals for cross-request identity or elicitation continuity.
Minimal server and View
import { MCPServer } from "mcp-use";
import { z } from "zod";
const server = new MCPServer({ name: "catalog", version: "1.0.0" });
export const showProduct = server.tool(
{
name: "show-product",
description: "Show one catalog product",
inputSchema: z.object({ id: z.string() }),
outputSchema: z.object({ id: z.string(), name: z.string() }),
view: { name: "product" },
},
async ({ id }) => {
const product = { id, name: "Example product" };
return {
content: [{ type: "text", text: JSON.stringify(product) }],
structuredContent: product,
};
},
);
export default server;
// views/product/view.tsx
import { ThemeProvider, useToolContext } from "mcp-use/react";
export default function ProductView() {
const view = useToolContext<"show-product">();
if (view.status === "pending") return <p>Loading…</p>;
if (view.status === "error") return <p>{view.error.message}</p>;
return <ThemeProvider>{view.toolOutput.name}</ThemeProvider>;
}
Guardrails
- Do not copy examples from v1 docs or historical changelogs.
- Do not invent exports or configuration fields; confirm them in installed declarations or source.
- Do not return a plain domain object from a tool callback.
- Do not bind a View without an
outputSchemaand matchingstructuredContent. - Do not claim success from a source build alone when package exports or interactive behavior changed.
- Do not deploy or mutate external systems unless the user explicitly requests it.
Run node <skill-dir>/scripts/check-v2.mjs <project-root> during migrations and reviews, then complete the focused checks in Verification.
Agent Skills
Put reusable agent workflows in skills/<name>/SKILL.md; the directory is
served automatically, so normally omit the skills server option. Use
skills: false to disable it or skills: { directory: "server-skills" } to
override the project-relative directory. Keep supporting references, scripts,
templates, and assets in the skill instead of inflating tool descriptions.