diff --git a/plugins/mcp-server-dev/skills/build-mcp-app/SKILL.md b/plugins/mcp-server-dev/skills/build-mcp-app/SKILL.md index 0ee1dd0..3038146 100644 --- a/plugins/mcp-server-dev/skills/build-mcp-app/SKILL.md +++ b/plugins/mcp-server-dev/skills/build-mcp-app/SKILL.md @@ -350,3 +350,4 @@ The `sleep` keeps stdin open long enough to collect all responses. Parse the jso - `references/iframe-sandbox.md` — CSP/sandbox constraints, the bundle-inlining pattern, image handling - `references/widget-templates.md` — reusable HTML scaffolds for picker / confirm / progress / display - `references/apps-sdk-messages.md` — the `App` class API: widget ↔ host ↔ server messaging +- `references/app-architecture-patterns.md` — multi-view app patterns: action dispatch, model context, build pipeline, dual transport diff --git a/plugins/mcp-server-dev/skills/build-mcp-app/references/app-architecture-patterns.md b/plugins/mcp-server-dev/skills/build-mcp-app/references/app-architecture-patterns.md new file mode 100644 index 0000000..c257781 --- /dev/null +++ b/plugins/mcp-server-dev/skills/build-mcp-app/references/app-architecture-patterns.md @@ -0,0 +1,381 @@ +# Multi-View App Architecture + +When a single-purpose widget isn't enough — when you have 5+ UI tools that share styling, state patterns, or a data model — consider building a **multi-view MCP app**: one React (or framework) SPA that dispatches to the right view based on which tool was called. + +This doc covers the production patterns for that architecture. It assumes you've read the main `build-mcp-app` skill and the `widget-templates.md` reference. + +> **Reference implementation:** [`nashville-charts-app`](https://github.com/bryankthompson/nashville-charts-app) on GitHub / npm demonstrates every pattern below. Install with `npx nashville-charts-app --stdio`. + +--- + +## Single resource, action-dispatched views + +Instead of one HTML file per tool, register **one shared resource** and point all UI tools at it. Each tool returns JSON with an `action` field that tells the app which view to render. + +**Server side:** + +```typescript +const RESOURCE_URI = "ui://my-app/app.html"; + +// Tool A — returns action + data +registerAppTool(server, "show_dashboard", { + description: "Show the analytics dashboard", + inputSchema: { range: z.enum(["week", "month"]) }, + _meta: { ui: { resourceUri: RESOURCE_URI } }, +}, async ({ range }) => { + const stats = await fetchStats(range); + return { + content: [{ type: "text", text: JSON.stringify({ action: "dashboard", stats }) }], + }; +}); + +// Tool B — same resource, different action +registerAppTool(server, "show_details", { + description: "Show detail view for a specific item", + inputSchema: { id: z.string() }, + _meta: { ui: { resourceUri: RESOURCE_URI } }, +}, async ({ id }) => { + const item = await fetchItem(id); + return { + content: [{ type: "text", text: JSON.stringify({ action: "detail", item }) }], + }; +}); + +// One resource serves all views +registerAppResource(server, "App", RESOURCE_URI, {}, + async () => ({ + contents: [{ uri: RESOURCE_URI, mimeType: RESOURCE_MIME_TYPE, text: appHtml }], + }), +); +``` + +**Widget side (React):** + +```tsx +function App() { + const [view, setView] = useState(null); + const { app } = useApp({ + appInfo: { name: "MyApp", version: "1.0.0" }, + onAppCreated: (app) => { + app.ontoolresult = async (result) => { + const data = JSON.parse(result.content[0].text); + if (data.action) setView(data); + }; + }, + }); + + if (!view) return
Waiting for tool call...
; + + switch (view.action) { + case "dashboard": return ; + case "detail": return ; + default: return
Unknown view
; + } +} +``` + +**Why this beats multiple HTML files:** +- Shared CSS, shared components, shared state (theme, loading indicators) +- One build artifact to deploy and cache +- Consistent UX across all views — users don't see a flash between different iframes +- The dispatch is just a `switch` on a string — trivial to extend + +--- + +## Model context updates + +`updateModelContext()` tells the LLM what the user is currently looking at — without adding a visible message to the chat. This is critical for multi-step workflows where Claude needs to reason about the current UI state. + +**What to send:** Structured state, not a prose description. YAML or JSON works well. + +```tsx +useEffect(() => { + if (!view || !app) return; + + const ctx = { + view: view.action, + ...(view.action === "dashboard" && { + range: view.stats.range, + itemCount: view.stats.items.length, + }), + ...(view.action === "detail" && { + itemId: view.item.id, + itemName: view.item.name, + }), + }; + + app.updateModelContext({ + content: [{ type: "text", text: JSON.stringify(ctx) }], + }).catch(() => { + // Host may not support updateModelContext — degrade silently + }); +}, [app, view]); +``` + +**When to send:** On every view change. Don't send on every keystroke or scroll — just when the semantic state changes (new view, new data, user selection). + +**Always catch:** Not all hosts support `updateModelContext()`. Wrap in `.catch()` so the app doesn't break in hosts that don't implement it. + +--- + +## Teardown guards + +After the host calls `onteardown`, other callbacks (`ontoolresult`, `onhostcontextchanged`) may still fire. Without a guard, these late callbacks can cause React state updates on an unmounted component. + +```tsx +const tornDown = useRef(false); + +app.onteardown = async () => { + tornDown.current = true; + setView(null); + return {}; +}; + +app.ontoolresult = async (result) => { + if (tornDown.current) return; // Guard + const data = JSON.parse(result.content[0].text); + if (data.action) setView(data); +}; + +app.onhostcontextchanged = () => { + if (tornDown.current) return; // Guard + setHostContext(app.getHostContext()); +}; +``` + +Use a `useRef` (not `useState`) — refs update synchronously and don't trigger re-renders. + +--- + +## Bidirectional tool calls from components + +When a UI component needs to call a server tool (e.g., a "Transpose" button, a "Load More" link), use `app.callServerTool()` and flow the result back through the parent's dispatch. + +```tsx +// Parent holds the dispatch +function App() { + const handleToolResult = useCallback((result) => { + const data = JSON.parse(result.content[0].text); + if (data.action) setView(data); + }, []); + + return ; +} + +// Child calls tools, parent re-dispatches +function DetailView({ item, app, onToolResult }) { + const [loading, setLoading] = useState(false); + + const handleRefresh = async () => { + setLoading(true); + try { + const result = await app.callServerTool({ + name: "show_details", + arguments: { id: item.id }, + }); + onToolResult(result); + } finally { + setLoading(false); + } + }; + + return ( +
+

{item.name}

+ +
+ ); +} +``` + +**Pattern:** Components don't set the view directly — they call tools and the parent handles the result. This keeps the data flow unidirectional: tool result -> parent dispatch -> child render. + +**App-only tools:** For tools that should only be callable from the widget (not by the LLM), use `server.registerTool` (not `registerAppTool` — these tools don't render a UI resource) with `visibility: ["app"]`: + +```typescript +server.registerTool("internal_action", { + description: "...", + inputSchema: { ... }, + _meta: { ui: { visibility: ["app"] } }, +}, handler); +``` + +--- + +## Build pipeline: Vite + esbuild + +A multi-view app needs a framework build (React, Vue, etc.) bundled into a single HTML file, plus a separate server bundle. Two-step build: + +**Step 1 — Vite bundles the SPA into one HTML file:** + +```typescript +// vite.config.ts +import { defineConfig } from "vite"; +import react from "@vitejs/plugin-react"; +import { viteSingleFile } from "vite-plugin-singlefile"; + +export default defineConfig({ + plugins: [react(), viteSingleFile()], + build: { + rollupOptions: { input: process.env.INPUT }, // e.g., INPUT=app.html + outDir: "dist", + emptyOutDir: false, // Don't delete dist — esbuild writes here too + }, +}); +``` + +`vite-plugin-singlefile` inlines all JS and CSS into the HTML. The output is a self-contained `dist/app.html` — no external chunks, no asset files. + +**Step 2 — esbuild bundles the server:** + +```bash +# Server module +npx esbuild server/server.ts --bundle --platform=node --format=esm \ + --outfile=dist/server.js --packages=external + +# CLI entry point (with shebang for npx) +npx esbuild main.ts --bundle --platform=node --format=esm \ + --outfile=dist/index.js --packages=external \ + --banner:js='#!/usr/bin/env node' +``` + +`--packages=external` keeps npm dependencies as imports (not bundled). The shebang banner makes `dist/index.js` directly executable. + +**Combined build script:** + +```json +{ + "build": "cross-env INPUT=app.html vite build && npx esbuild server/server.ts --bundle --platform=node --format=esm --outfile=dist/server.js --packages=external && npx esbuild main.ts --bundle --platform=node --format=esm --outfile=dist/index.js --packages=external --banner:js='#!/usr/bin/env node'" +} +``` + +**Path resolution (dev vs. prod):** + +```typescript +// Works in both TypeScript (dev) and compiled JS (prod) +const DIST_DIR = import.meta.filename.endsWith(".ts") + ? path.join(import.meta.dirname, "..", "dist") // Running .ts directly + : import.meta.dirname; // Running compiled .js + +const appHtml = await fs.promises.readFile( + path.join(DIST_DIR, "app.html"), "utf-8" +); +``` + +--- + +## Dual transport: HTTP + stdio + +Support both transports from one codebase. Use a server factory function and a CLI flag. + +```typescript +// main.ts +import { createServer } from "./server/server.js"; + +const useStdio = process.argv.includes("--stdio"); + +if (useStdio) { + startStdio(createServer); +} else { + startHTTP(createServer); +} +``` + +**HTTP (stateless, one server per request):** + +```typescript +async function startHTTP(createServer) { + const app = express(); + app.use(express.json()); + + app.all("/mcp", async (req, res) => { + const server = createServer(); // Fresh per request + const transport = new StreamableHTTPServerTransport({ + sessionIdGenerator: undefined, // Stateless + }); + res.on("close", () => { + transport.close(); + server.close(); + }); + await server.connect(transport); + await transport.handleRequest(req, res, req.body); + }); + + app.listen(process.env.PORT ?? 3001); +} +``` + +**stdio (persistent, one server for the session):** + +```typescript +async function startStdio(createServer) { + const server = createServer(); + const transport = new StdioServerTransport(); + await server.connect(transport); + // Server stays alive until the process exits +} +``` + +**Graceful shutdown** for HTTP: + +```typescript +const shutdown = () => { + httpServer.close(() => process.exit(0)); +}; +process.on("SIGINT", shutdown); +process.on("SIGTERM", shutdown); +``` + +--- + +## Debugging tips + +**Keep stdout clean.** In stdio mode, stdout IS the MCP protocol. All debug output must go to stderr: + +```typescript +console.error("[DEBUG] Tool called:", toolName); // stderr — safe +console.log("..."); // stdout — breaks protocol +``` + +**Extension detection.** The MCP SDK's Zod schema strips unknown fields from `capabilities` — including `extensions`, which signals UI support. To check if the client supports your widgets, intercept raw messages before SDK parsing: + +```typescript +transport.onmessage = (msg) => { + if (msg?.method === "initialize") { + const extensions = msg.params?.capabilities?.extensions; + if (extensions) { + console.error("[DEBUG] Client supports extensions:", JSON.stringify(extensions)); + } + } +}; +``` + +Compare raw messages with SDK-parsed `server.getClientCapabilities()` to diagnose stripping. + +**Resource caching in Claude Desktop.** After editing widget HTML, fully quit (Cmd+Q) and relaunch. Window-close doesn't clear the resource cache. + +--- + +## Prompt templates as workflow orchestration + +MCP prompts can guide the LLM through multi-tool workflows — acting as lightweight choreography: + +```typescript +server.registerPrompt("analyze_pipeline", { + title: "Run Full Analysis", + description: "Fetch data, visualize, and summarize", +}, () => ({ + messages: [{ + role: "user", + content: { + type: "text", + text: "Run a full analysis: use fetch-data to pull the latest numbers, " + + "show-dashboard to visualize them, then summarize the key insights.", + }, + }], +})); +``` + +The prompt tells the LLM which tools to call in which order. No dynamic data — just instructions. This surfaces as a slash-command in hosts that support prompts, giving users a one-click workflow trigger.