The Model Context Protocol (MCP) has become a cornerstone of how AI assistants (Cursor, Claude Desktop, Claude Code, Zed, Windsurf) discover, read, and install Sora UI components.

When we first built apps/mcp, the MCP ecosystem was young. We leaned on mcp-framework, wrapping it with custom Vercel serverless adapters, middleware path rewrites, and bespoke build scripts. While it served us well initially, maintaining that glue code became friction: local STDIO development required manual setups, adding new MCP primitives (like Prompts and Resources) wasn't seamless, and bundle outputs were tied to complex serverless configurations.

Today, we have completely rebuilt and migrated our MCP server into apps/xmcp, powered by the modern xmcp framework and @xmcp-dev/compiler.

Here is a full breakdown of why we migrated, what changed under the hood, and how the old and new architectures compare.

Old (apps/mcp) vs New (apps/xmcp)

Why We Chose xmcp

1. First-Class Dual Bundling (HTTP + STDIO)

AI clients connect to MCP servers in different ways:

• Remote clients & team setups connect over Streamable HTTP / Server-Sent Events (SSE).
• Local desktop tools (like Claude Desktop or local Cursor instances) perform best over STDIO without network overhead.

With the old apps/mcp, supporting both meant maintaining separate entry points and fighting serverless bundling quirks.

xmcp solves this natively. Running bun run build in apps/xmcp compiles two optimized standalone bundles:

• dist/http.js: Streamable HTTP server with built-in CORS, health checks, and route handling.
• dist/stdio.js: Zero-dependency STDIO bridge that streams JSON-RPC directly to stdin/stdout.

cd apps/xmcp
bun install
bun run build    # produces dist/http.js and dist/stdio.js

2. Standardized /mcp Streaming Endpoint

In the previous version, JSON-RPC requests hit the root path /, causing conflicts when trying to host a human-readable landing page or discovery metadata on the same domain.

With apps/xmcp, the live MCP streaming endpoint is cleanly located at:

https://mcp.soralabs.io.vn/mcp

Visiting the root domain (https://mcp.soralabs.io.vn) now serves a clean visual dashboard, while AI clients connect directly to /mcp.

Expanding MCP Capabilities: Prompts & Resources

In addition to tools, the new server takes full advantage of the Model Context Protocol specification:

apps/xmcp/src/
├── tools/        — search_docs, get_page, list_sections, get_component_info
├── prompts/      — install-component
└── resources/    — registry-catalog

1. Prompts (src/prompts/install-component.ts)

Instead of requiring users to remember the exact syntax for component installation, apps/xmcp exposes the install-component prompt template. When invoked, it orchestrates the entire agent flow:

1. Fetches component metadata via get_component_info.
2. Emits the safe, non-interactive install command (npx shadcn@latest add @soralabs/<name> --yes).
3. Handles workspace resolution via --cwd if working in a monorepo.
4. Verifies the installed files and explains how to integrate the component.

2. Resources (src/resources/registry.ts)

AI assistants can now attach the registry-catalog resource directly into their context window. It provides a real-time, machine-readable JSON snapshot of all installable UI components, animation primitives, and hooks without multiple iterative search queries.

3. Refined Tools Suite (src/tools/)

All four core tools have been upgraded with tighter schemas and smarter token management:

• search_docs: Fast Orama search with fallback parsing, section filtering (documentation, components, catalog, motion, ui), capped at 25 results.
• get_page: Fetches clean markdown documentation with a strict 8,000-token budget and automatic truncation notices.
• list_sections: Visual tree outline of all available documentation sections and page counts parsed from llms.txt.
• get_component_info: Intelligently returns non-interactive installation lines (npx shadcn@latest add @soralabs/<name> --yes and sora-cli), dependency lists, and opt-in source code (includeSource: true).

How to Connect to the New Server

Updating your configuration is straightforward.

Cursor (.cursor/mcp.json)

Remote HTTP (Recommended):

{
  "mcpServers": {
    "sora-ui": {
      "url": "https://mcp.soralabs.io.vn/mcp"
    }
  }
}

Local STDIO (from source):

{
  "mcpServers": {
    "sora-ui": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/sora/apps/xmcp/dist/stdio.js"]
    }
  }
}

Claude Code

# Connect via remote HTTP
claude mcp add --transport http sora-ui https://mcp.soralabs.io.vn/mcp

# Or connect via local STDIO
claude mcp add sora-ui node /ABSOLUTE/PATH/TO/sora/apps/xmcp/dist/stdio.js

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "sora-ui": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.soralabs.io.vn/mcp"]
    }
  }
}

Next Steps

With apps/xmcp in place and deployed, we will be deprecating and removing the legacy apps/mcp directory. All documentation pages, client configurations, and links have been updated to point to the new /mcp endpoint.

For full setup guides and interactive install links, visit the updated MCP Documentation.