MCPInstall

build-mcp-app

This skill should be used when the user wants to build an "MCP app", add "interactive UI" or "widgets" to an MCP server, "render components in chat", build "MCP UI resources", make a tool that shows a "form", "picker", "dashboard" or "confirmation dialog" inline in the conversation, or mentions "apps SDK" in the context of MCP. Use AFTER the build-mcp-server skill has settled the deployment model, or when the user already knows they want UI widgets.

Installs
6.6K
Last 24h
+48
Trending rank
#3440
Security
Pass

Install

Works with every agent the skills CLI supports.

npx skills add https://github.com/anthropics/claude-plugins-official --skill build-mcp-app
  1. 01

    Install

    Run the command in your project. The CLI asks which agents to install for.

  2. 02

    Describe the task

    Ask your agent for something this skill covers. It loads build-mcp-app automatically when relevant.

  3. 03

    Review the result

    The agent follows the SKILL.md steps, scripts and references bundled with the skill.

Security audits

Independent scans of the skill's instructions and bundled files.

Gen Agent Trust Hub PassSAFE ·This skill teaches AI agents how to build interactive Model Context Protocol (MCP) applications with embedded UI elements. It does not introduce code patterns that pose a risk or compromise safety guidelines. All recommended practices align with proper software development and standard platform controls.
Socket PassNo alerts
Snyk PassLOW ·Risk: LOW · No issues
ZeroLeaks PassNONE ·Score: 93/100 · 2 sections analyzed

SKILL.md

The instructions your agent reads when the skill is triggered.

Build an MCP App (Interactive UI Widgets)

An MCP app is a standard MCP server that also serves UI resources — interactive components rendered inline in the chat surface. Build once, runs in Claude and ChatGPT and any other host that implements the apps surface.

The UI layer is additive. Under the hood it's still tools, resources, and the same wire protocol. If you haven't built a plain MCP server before, the build-mcp-server skill covers the base layer. This skill adds widgets on top.

Testing in Claude: Add the server as a custom connector in claude.ai (via a Cloudflare tunnel for local dev) — this exercises the real iframe sandbox and hostContext. See https://claude.com/docs/connectors/building/testing.

Claude host specifics

_meta.ui.* keyWhereEffect
resourceUritoolWhich ui:// resource the host renders for this tool's results.
visibility: ["app"]toolHide a widget-only helper tool (e.g. geometry/image fetcher called via callServerTool) from Claude's tool list.
prefersBorder: falseresourceDrop the host's outer card border (mobile).
csp.{connectDomains, resourceDomains, baseUriDomains}resourceDeclare external origins; default is block-all. frameDomains is currently restricted in Claude.
  • hostContext.safeAreaInsets: {top, right, bottom, left} (px) — honor these for notches and the composer overlay.
  • Directory submission requires OAuth or authless (none) — static bearer is private-deploy only and blocks listing — plus tool annotations and 3–5 PNG screenshots; see references/directory-checklist.md.

When a widget beats plain text

Don't add UI for its own sake — most tools are fine returning text or JSON. Add a widget when one of these is true:

SignalWidget type
Tool needs structured input Claude can't reliably inferForm
User must pick from a list Claude can't rank (files, contacts, records)Picker / table
Destructive or billable action needs explicit confirmationConfirm dialog
Output is spatial or visual (charts, maps, diffs, previews)Display widget
Long-running job the user wants to watchProgress / live status

If none apply, skip the widget. Text is faster to build and faster for the user.


Widgets vs Elicitation — route correctly

Before building a widget, check if elicitation covers it. Elicitation is spec-native, zero UI code, works in any compliant host.

NeedElicitationWidget
Confirm yes/no✅overkill
Pick from short enum✅overkill
Fill a flat form (name, email, date)✅overkill
Pick from a large/searchable list❌ (no scroll/search)✅
Visual preview before choosing❌✅
Chart / map / diff view❌✅
Live-updating progress❌✅

If elicitation covers it, use it. See ../build-mcp-server/references/elicitation.md.


Architecture: two deployment shapes

Remote MCP app (most common)

Hosted streamable-HTTP server. Widget templates are served as resources; tool results reference them. The host fetches the resource, renders it in an iframe sandbox, and brokers messages between the widget and Claude.

┌──────────┐  tools/call   ┌────────────┐
│  Claude  │─────────────> │ MCP server │
│   host   │<── result ────│  (remote)  │
│          │  + widget ref │            │
│          │               │            │
│          │ resources/read│            │
│          │─────────────> │  widget    │
│ ┌──────┐ │<── template ──│  HTML/JS   │
│ │iframe│ │               └────────────┘
│ │widget│ │
│ └──────┘ │
└──────────┘

MCPB-packaged MCP app (local + UI)

Same widget mechanism, but the server runs locally inside an MCPB bundle. Use this when the widget needs to drive a local application — e.g., a file picker that browses the actual local disk, a dialog that controls a desktop app.

For MCPB packaging mechanics, defer to the build-mcpb skill. Everything below applies to both shapes.


How widgets attach to tools

A widget-enabled tool has two separate registrations:

  1. The tool declares a UI resource via _meta.ui.resourceUri. Its handler returns plain text/JSON — NOT the HTML.
  2. The resource is registered separately and serves the HTML.

When Claude calls the tool, the host sees _meta.ui.resourceUri, fetches that resource, renders it in an iframe, and pipes the tool's return value into the iframe via the ontoolresult event.

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { registerAppTool, registerAppResource, RESOURCE_MIME_TYPE }
  from "@modelcontextprotocol/ext-apps/server";
import { z } from "zod";

const server = new McpServer({ name: "contacts", version: "1.0.0" });

// 1. The tool — returns DATA, declares which UI to show
registerAppTool(server, "pick_contact", {
  description: "Open an interactive contact picker",
  annotations: { title: "Pick Contact", readOnlyHint: true },
  inputSchema: { filter: z.string().optional() },
  _meta: { ui: { resourceUri: "ui://widgets/contact-picker.html" } },
}, async ({ filter }) => {
  const contacts = await db.contacts.search(filter);
  // Plain JSON — the widget receives this via ontoolresult
  return { content: [{ type: "text", text: JSON.stringify(contacts) }] };
});

// 2. The resource — serves the HTML
registerAppResource(
  server,
  "Contact Picker",
  "ui://widgets/contact-picker.html",
  {},
  async () => ({
    contents: [{
      uri: "ui://widgets/contact-picker.html",
      mimeType: RESOURCE_MIME_TYPE,
      text: pickerHtml,  // your HTML string
    }],
  }),
);

The URI scheme ui:// is convention. The mime type MUST be RESOURCE_MIME_TYPE ("text/html;profile=mcp-app") — this is how the host knows to render it as an interactive iframe, not just display the source.


Widget runtime — the App class

Inside the iframe, your script talks to the host via the App class from @modelcontextprotocol/ext-apps. This is a persistent bidirectional connection — the widget stays alive as long as the conversation is active, receiving new tool results and sending user actions.

<script type="module">
  /* ext-apps bundle inlined at build time → globalThis.ExtApps */
  /*__EXT_APPS_BUNDLE__*/
  const { App } = globalThis.ExtApps;

  const app = new App({ name: "ContactPicker", version: "1.0.0" }, {});

  // Set handlers BEFORE connecting
  app.ontoolresult = ({ content }) => {
    const contacts = JSON.parse(content[0].text);
    render(contacts);
  };

  await app.connect();

  // Later, when the user clicks something:
  function onPick(contact) {
    app.sendMessage({
      role: "user",
      content: [{ type: "text", text: `Selected contact: ${contact.id}` }],
    });
  }
</script>

The /*__EXT_APPS_BUNDLE__*/ placeholder gets replaced by the server at startup with the contents of @modelcontextprotocol/ext-apps/app-with-deps — see references/iframe-sandbox.md for why this is necessary and the rewrite snippet. Do not import { App } from "https://esm.sh/..."; the iframe's CSP blocks the transitive dependency fetches and the widget renders blank.

MethodDirectionUse for
app.ontoolresult = fnHost → widgetReceive the tool's return value
app.ontoolinput = fnHost → widgetReceive the tool's input args (what Claude passed)
app.sendMessage({...})Widget → hostInject a message into the conversation
app.updateModelContext({...})Widget → hostUpdate context silently (no visible message)
app.callServerTool({name, arguments})Widget → serverCall another tool on your server
app.openLink({url})Widget → hostOpen a URL in a new tab (sandbox blocks window.open)
app.getHostContext() / app.onhostcontextchangedHost → widgetTheme, host CSS vars, containerDimensions, displayMode, deviceCapabilities
app.requestDisplayMode({mode})Widget → hostAsk for inline / pip / fullscreen
app.downloadFile({name, mimeType, content})Widget → hostHost-mediated download (base64 content)
new App(info, caps, {autoResize: true})—Iframe height tracks rendered content

sendMessage is the typical "user picked something, tell Claude" path. updateModelContext is for state that Claude should know about but shouldn't clutter the chat. openLink is required for any outbound navigation — window.open and <a target="_blank"> are blocked by the sandbox attribute.

What widgets cannot do:

  • Access the host page's DOM, cookies, or storage
  • Make network calls to arbitrary origins (CSP-restricted — route through callServerTool)
  • Open popups or navigate directly — use app.openLink({url})
  • Load remote images reliably — inline as data: URLs server-side

Keep widgets small and single-purpose. A picker picks. A chart displays. Don't build a whole sub-app inside the iframe — split it into multiple tools with focused widgets.


Scaffold: minimal picker widget

Install:

npm install @modelcontextprotocol/sdk @modelcontextprotocol/ext-apps zod express

Server (src/server.ts):

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";
import { registerAppTool, registerAppResource, RESOURCE_MIME_TYPE }
  from "@modelcontextprotocol/ext-apps/server";
import express from "express";
import { readFileSync } from "node:fs";
import { createRequire } from "node:module";
import { z } from "zod";

const require = createRequire(import.meta.url);
const server = new McpServer({ name: "contact-picker", version: "1.0.0" });

// Inline the ext-apps browser bundle into the widget HTML.
// The iframe CSP blocks CDN script fetches — bundling is mandatory.
const bundle = readFileSync(
  require.resolve("@modelcontextprotocol/ext-apps/app-with-deps"), "utf8",
).replace(/export\{([^}]+)\};?\s*$/, (_, body) =>
  "globalThis.ExtApps={" +
  body.split(",").map((p) => {
    const [local, exported] = p.split(" as ").map((s) => s.trim());
    return `${exported ?? local}:${local}`;
  }).join(",") + "};",
);
const pickerHtml = readFileSync("./widgets/picker.html", "utf8")
  .replace("/*__EXT_APPS_BUNDLE__*/", () => bundle);

registerAppTool(server, "pick_contact", {
  description: "Open an interactive contact picker. User selects one contact.",
  annotations: { title: "Pick Contact", readOnlyHint: true },
  inputSchema: { filter: z.string().optional().describe("Name/email prefix filter") },
  _meta: { ui: { resourceUri: "ui://widgets/picker.html" } },
}, async ({ filter }) => {
  const contacts = await db.contacts.search(filter ?? "");
  return { content: [{ type: "text", text: JSON.stringify(contacts) }] };
});

registerAppResource(server, "Contact Picker", "ui://widgets/picker.html", {},
  async () => ({
    contents: [{ uri: "ui://widgets/picker.html", mimeType: RESOURCE_MIME_TYPE, text: pickerHtml }],
  }),
);

const app = express();
app.use(express.json());
app.post("/mcp", async (req, res) => {
  const transport = new StreamableHTTPServerTransport({ sessionIdGenerator: undefined });
  res.on("close", () => transport.close());
  await server.connect(transport);
  await transport.handleRequest(req, res, req.body);
});
app.listen(process.env.PORT ?? 3000);

For local-only widget apps (driving a desktop app, reading local files), swap the transport to StdioServerTransport and package via the build-mcpb skill.

Widget (widgets/picker.html):

<!doctype html>
<meta charset="utf-8" />
<style>
  body { font: 14px system-ui; margin: 0; }
  ul { list-style: none; padding: 0; margin: 0; max-height: 300px; overflow-y: auto; }
  li { padding: 10px 14px; cursor: pointer; border-bottom: 1px solid #eee; }
  li:hover { background: #f5f5f5; }
  .sub { color: #666; font-size: 12px; }
</style>
<ul id="list"></ul>
<script type="module">
/*__EXT_APPS_BUNDLE__*/
const { App } = globalThis.ExtApps;
(async () => {
  const app = new App({ name: "ContactPicker", version: "1.0.0" }, {});
  const ul = document.getElementById("list");

  app.ontoolresult = ({ content }) => {
    const contacts = JSON.parse(content[0].text);
    ul.innerHTML = "";
    for (const c of contacts) {
      const li = document.createElement("li");
      li.innerHTML = `<div>${c.name}</div><div class="sub">${c.email}</div>`;
      li.addEventListener("click", () => {
        app.sendMessage({
          role: "user",
          content: [{ type: "text", text: `Selected contact: ${c.id} (${c.name})` }],
        });
      });
      ul.append(li);
    }
  };

  await app.connect();
})();
</script>

See references/widget-templates.md for more widget shapes.


Design notes that save you a rewrite

One widget per tool. Resist the urge to build one mega-widget that does everything. One tool → one focused widget → one clear result shape. Claude reasons about these far better.

Tool description must mention the widget. Claude only sees the tool description when deciding what to call. "Opens an interactive picker" in the description is what makes Claude reach for it instead of guessing an ID.

Widgets are optional at runtime. Hosts that don't support the apps surface simply ignore _meta.ui and render the tool's text content normally. Since your tool handler already returns meaningful text/JSON (the widget's data), degradation is automatic — Claude sees the data directly instead of via the widget.

Don't block on widget results for read-only tools. A widget that just displays data (chart, preview) shouldn't require a user action to complete. Return the display widget and a text summary in the same result so Claude can continue reasoning without waiting.

Layout-fork by item count, not by tool count. If one use case is "show one result in detail" and another is "show many results side-by-side", don't make two tools — make one tool that accepts items[], and let the widget pick a layout: items.length === 1 → detail view, > 1 → carousel. Keeps the server schema simple and lets Claude decide count naturally.

Put Claude's reasoning in the payload. A short note field on each item (why Claude picked it) rendered as a callout on the card gives users the reasoning inline with the choice. Mention this field in the tool description so Claude populates it.

Normalize image shapes server-side. If your data source returns images with wildly varying aspect ratios, rewrite to a predictable variant (e.g. square-bounded) before fetching for the data-URL inline. Then give the widget's image container a fixed aspect-ratio + object-fit: contain so everything sits centered.

Follow host theme. app.getHostContext()?.theme (after connect()) plus app.onhostcontextchanged for live updates. Toggle a .dark class on <html>, keep colors in CSS custom props with a :root.dark {} override block, set color-scheme. Disable mix-blend-mode: multiply in dark — it makes images vanish.


Testing

Claude Desktop — current builds still require the command/args config shape (no native "type": "http"). Wrap with mcp-remote and force http-only transport so the SSE probe doesn't swallow widget-capability negotiation:

{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:3000/mcp",
               "--allow-http", "--transport", "http-only"]
    }
  }
}

Desktop caches UI resources aggressively. After editing widget HTML, fully quit (⌘Q / Alt+F4, not window-close) and relaunch to force a cold resource re-fetch.

Headless JSON-RPC loop — fast iteration without clicking through Desktop:

# test.jsonl — one JSON-RPC message per line
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"t","version":"0"}}}
{"jsonrpc":"2.0","method":"notifications/initialized"}
{"jsonrpc":"2.0","id":2,"method":"tools/list"}
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"your_tool","arguments":{...}}}

(cat test.jsonl; sleep 10) | npx mcp-remote http://localhost:3000/mcp --allow-http

The sleep keeps stdin open long enough to collect all responses. Parse the jsonl output with jq or a Python one-liner.

Widget dev loop — avoid the ⌘Q-relaunch cycle entirely by serving the inlined widget HTML at a plain GET route with a fake ExtApps shim that fires ontoolresult from a query param:

app.get("/widget-preview", (_req, res) => {
  const shim = `globalThis.ExtApps={applyHostStyleVariables:()=>{},App:class{
    constructor(){this.h={}} ontoolresult;onhostcontextchanged;
    async connect(){const p=new URLSearchParams(location.search).get("payload");
      if(p)this.ontoolresult?.({content:[{type:"text",text:p}]});}
    getHostContext(){return{theme:"light"}}
    sendMessage(m){console.log("sendMessage",m)} updateModelContext(){}
    callServerTool(){return Promise.resolve({content:[]})} openLink(){} downloadFile(){}
  }};`;
  res.type("html").send(widgetHtml.replace("/*__EXT_APPS_BUNDLE__*/", shim));
});

Open http://localhost:3000/widget-preview?payload={"rows":[...]} in a normal browser tab and iterate with ordinary devtools.

Host fallback — use a host without the apps surface (or MCP Inspector) and confirm the tool's text content degrades gracefully.

CSP debugging — open the iframe's own devtools console. CSP violations are the #1 reason widgets silently fail (blank rectangle, no error in the main console). See references/iframe-sandbox.md.


Reference files

  • references/iframe-sandbox.md — CSP/sandbox constraints, the bundle-inlining pattern, image handling, host theming
  • 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, lifecycle & supersession
  • references/payload-budgeting.md — host tool-result size caps, prune-then-truncate, heavy assets via callServerTool
  • references/abuse-protection.md — Anthropic egress CIDRs, tiered rate limiting, trust proxy, response caching
  • references/directory-checklist.md — pre-flight for connector-directory submission

FAQ

What does the build-mcp-app skill do?

This skill should be used when the user wants to build an "MCP app", add "interactive UI" or "widgets" to an MCP server, "render components in chat", build "MCP UI resources", make a tool that shows a "form", "picker", "dashboard" or "confirmation dialog" inline in the conversation, or mentions "apps SDK" in the context of MCP. Use AFTER the build-mcp-server skill has settled the deployment model, or when the user already knows they want UI widgets.

How do I install build-mcp-app?

Run npx skills add https://github.com/anthropics/claude-plugins-official --skill build-mcp-app in your project. The skills CLI detects the agents you use and copies the skill into the right folder. For Claude Code you can also place the skill folder at ~/.claude/skills/build-mcp-app/ (personal) or .claude/skills/build-mcp-app/ (project).

Which AI agents can use build-mcp-app?

Any agent that supports the open Agent Skills format (a folder with a SKILL.md file), including Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf, Cline, OpenCode and more.

Is build-mcp-app safe to install?

Security partners have reviewed it: Gen Agent Trust Hub passed; Socket passed; Snyk passed; ZeroLeaks passed. As with any skill, read the SKILL.md and bundled scripts before installing.

Who maintains build-mcp-app?

It is published by anthropics in anthropics/claude-plugins-official, and it is an official skill from the maker of the technology it covers. It has 6,555 recorded installs.

More from anthropics/claude-plugins-official

View all →

frontend-design

anthropics/claude-plugins-official

Guidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.

OfficialFrontend & Design17.1K installs

claude-md-improver

anthropics/claude-plugins-official

Audit and improve CLAUDE.md files in repositories. Use when user asks to check, audit, update, improve, or fix CLAUDE.md files. Scans for all CLAUDE.md files, evaluates quality against templates, outputs quality report, then makes targeted updates. Also use when the user mentions "CLAUDE.md maintenance" or "project memory optimization".

OfficialSecurity12.8K installs

claude-automation-recommender

anthropics/claude-plugins-official

Analyze a codebase and recommend Claude Code automations (hooks, subagents, skills, plugins, MCP servers). Use when user asks for automation recommendations, wants to optimize their Claude Code setup, mentions improving Claude Code workflows, asks how to first set up Claude Code for a project, or wants to know what Claude Code features they should use.

OfficialProductivity & Automation9.3K installs

playground

anthropics/claude-plugins-official

Creates interactive HTML playgrounds — self-contained single-file explorers that let users configure something visually through controls, see a live preview, and copy out a prompt. Use when the user asks to make a playground, explorer, or interactive tool for a topic.

OfficialFrontend & Design8.1K installs

skill-creator

anthropics/claude-plugins-official

Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.

OfficialAI & Agents7.7K installs

agent-development

anthropics/claude-plugins-official

This skill should be used when the user asks to "create an agent", "add an agent", "write a subagent", "agent frontmatter", "when to use description", "agent examples", "agent tools", "agent colors", "autonomous agent", or needs guidance on agent structure, system prompts, triggering conditions, or agent development best practices for Claude Code plugins.

OfficialAI & Agents7.4K installs

Popular ai & agents skills

View category →

google-agents-cli-publish

google/agents-cli

This skill should be used when the user wants to "publish an agent", "publish my ADK agent", "register an agent with Gemini", "publish to Gemini" (formerly Gemini Enterprise), or needs guidance on the agents-cli publish gemini command. Also use when the user wants to "manage agents in Agent Registry", "list/update/delete registered agents", or "register an MCP server". Covers ADK vs A2A registration modes, programmatic and interactive usage, flag reference, auto-detection from deployment metadata, Agent Registry fleet management, and troubleshooting. Part of the agents-cli skills suite. Do NOT use for deployment (use google-agents-cli-deploy).

AI & Agents513.9K installs

google-agents-cli-eval

google/agents-cli

This skill should be used when the user wants to "run an evaluation", "evaluate my agent", "evaluate my ADK agent", "write an eval dataset", "analyze eval failures", "compare eval results", "optimize agent", or needs guidance on the Gemini platform eval methodology and the Quality Flywheel. Covers eval metrics, dataset schema, LLM-as-judge scoring, and common failure causes. Applies to any agents-cli project, whatever framework the agent is written in. Do NOT use for agent API code patterns (ADK: use google-agents-cli-adk-code), deployment (use google-agents-cli-deploy), or project scaffolding (use google-agents-cli-scaffold).

AI & Agents513.5K installs

skill-creator

anthropics/skills

Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.

OfficialAI & Agents416.4K installs

codex-pet

prime-skills/runcomfy-agent-skills

Codex Pet generator on RunComfy. Build a Codex-compatible Codex Pet spritesheet.webp + pet.json from a single reference image, drop it into ${CODEXHOME:-$HOME/.codex}/pets/ / and Codex picks it up as a custom Codex Pet next to the 8 built-ins. This skill produces the exact Codex Pet atlas Codex expects (1536x1872 PNG/WebP, 8 cols x 9 rows, 192x208 cells, 9 animation states — idle, running-right, running-left, waving, jumping, failed, waiting, running, review). Calls OpenAI GPT Image 2 edit ONCE via the local RunComfy CLI as runcomfy run openai/gpt-image-2/edit to produce a canonical Codex Pet pose, then assembles all 9 animation rows programmatically with ImageMagick micro-transforms — no Codex Pro, no $imagegen, no OPENAIAPIKEY required, only RUNCOMFYTOKEN. Triggers on "codex pet", "create codex pet", "make codex pet", "hatch codex pet", "/hatch image", "desktop pet codex", "codex pets", "spritesheet.webp", or any explicit ask to build a custom pet for OpenAI Codex.

AI & Agents402.1K installs

full-output-enforcement

leonxlnx/taste-skill

Overrides default LLM truncation behavior. Enforces complete code generation, bans placeholder patterns, and handles token-limit splits cleanly. Apply to any task requiring exhaustive, unabridged output.

AI & Agents401K installs

executing-plans

obra/superpowers

Use when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available

AI & Agents245.7K installs