Building AI-operated in-app product experiences using remote MCP servers

Model Context Protocol (MCP) is usually framed as a way for LLMs to read local files or query SQL databases. But the highest-leverage application of MCP isn't reading data—it's runtime app orchestration.

What if your AI coding agent could ship in-app announcements, contextual onboarding guides, and feature callouts directly from your terminal without writing ephemeral UI code?

---

The Problem with LLMs Generating Frontend Code on the Fly

Having an AI agent write ad-hoc React components for temporary announcements is risky:

• It introduces bundle bloat and dead components.
• CSS inconsistencies quickly creep into your design system.
• It risks breaking page layouts during SSR hydration.

The better architectural pattern is Deterministic JSON Rules governed by MCP.

How the Remote MCP Flow Works

Here is the architecture:

1. Agent Tool Call: You prompt Claude Code or Cursor to announce a release.
2. MCP Remote Server: The agent connects to https://www.neotic.app/api/mcp using OAuth 2.1 PKCE S256.
3. Rule Specification: The agent emits a structured payload declaring the trigger conditions:

{
  "slot_id": "onboarding-welcome-modal",
  "rules": {
    "path_pattern": "/onboarding/*",
    "user_properties": {
      "account_age_days": { "lte": 3 },
      "completed_setup": false
    }
  },
  "content": {
    "headline": "Welcome to the new dashboard",
    "cta": { "label": "Take 1-min Tour", "url": "/tour" }
  }
}

1. Client Evaluation: The browser evaluates these rules locally using a micro-SDK (<4KB) with zero server latency.

Connecting MCP to Your Claude / Cursor Config

You can configure this in your mcp.json or Claude desktop configuration:

{
  "mcpServers": {
    "neotic": {
      "url": "https://www.neotic.app/api/mcp",
      "transport": "sse"
    }
  }
}

Check out Neotic's Developer Docs to see the full MCP specification and test out remote triggers.

---

💬 Let's Discuss

How does your team currently manage in-app announcements and feature onboarding? Do you hardcode modals into React, use an external script, or automate via MCP? Drop your thoughts below!

(If you're building with Next.js or AI coding agents, check out Neotic and our remote MCP endpoint at https://www.neotic.app/api/mcp!)