MCP (Model Context Protocol)

What is MCP?

MCP (Model Context Protocol) is a specification published by Anthropic in late 2024 and adopted across the AI industry through 2025-2026. It defines a standard wire format for AI clients (Claude Desktop, Claude Code, Cursor, ChatGPT Pro Connectors, Zed, Google Antigravity, n8n's AI Agent node) to discover and invoke external tools without per-vendor custom integration. Before MCP, every AI app shipped its own 'function calling' layer with proprietary schemas; MCP unifies that into one open protocol.

At a technical level: an MCP server is a process (typically run via stdio) that exposes a JSON-RPC interface. The client connects, calls list_tools to get the catalog, sees tool names, descriptions, and JSON Schema input definitions, then calls call_tool with structured arguments. The server executes the tool and returns structured output. The model picks tools to call based on the user's natural-language prompt; the protocol handles the plumbing.

Why MCP matters for SaaS products

Pre-MCP: a SaaS that wanted AI agent integration had to build separate adapters for each client (a Claude integration, a Cursor integration, a custom OpenAI function-calling wrapper, etc.). Each adapter required vendor-specific code, vendor-specific authentication, vendor-specific deployment. Post-MCP: build one MCP server. Every compatible AI client gets the integration for free. New clients (Antigravity, Zed) light up automatically the day they ship MCP support.

For users this means natural-language access to SaaS features without switching apps. 'Schedule my new product video to TikTok and Instagram for tomorrow 9am EST' executes against your social scheduler from inside Claude Desktop, with no separate UI to learn. The agent picks the right tool (schedule_post), fills in arguments (scheduled_date, platforms array, media_urls), and dispatches.

How an MCP integration looks in practice

Three steps. First, the SaaS publishes an MCP server (commonly as an npm or PyPI package). For CodivUpload, that's `npx codivupload-mcp` — a single command that starts the server. Second, the user adds the server to their AI client config. For Claude Desktop, that's a JSON entry in claude_desktop_config.json with command: 'npx', args: ['codivupload-mcp'], and an env block with the API key. Third, the user prompts the AI in plain language. The AI lists available tools (publish_post, schedule_post, list_profiles, get_analytics, etc.), picks the right one, fills arguments from context, and calls it.

The AI never sees the API key — it lives in the env block of the local MCP server config and never crosses the wire to the model provider. This is a meaningful security property: the AI does the reasoning, the local MCP server holds the credentials.

MCP vs traditional REST API

Both exist for the same purpose (programmatic access to a SaaS). The differences: REST API needs glue code per client (SDK calls, auth handling, response parsing); MCP needs a one-time config block. REST API is for when you write the orchestration; MCP is for when an AI does. REST API supports any caller; MCP is specifically for AI agents. Most modern SaaS products with serious AI strategy ship both — REST for direct integration, MCP for AI clients. CodivUpload publishes both: api.codivupload.com (REST) and npx codivupload-mcp (MCP).

Connect Claude Desktop to CodivUpload's MCP server

json

// ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
// %APPDATA%/Claude/claude_desktop_config.json (Windows)
{
  "mcpServers": {
    "codivupload": {
      "command": "npx",
      "args": ["-y", "codivupload-mcp"],
      "env": {
        "CODIVUPLOAD_API_KEY": "cdv_..."
      }
    }
  }
}

// Then in Claude:
// > Schedule my new product video to TikTok, Instagram, and YouTube
// > for tomorrow 9am EST. Use the launch caption I drafted in the doc.
//
// Claude lists tools, picks schedule_post, fills arguments, and calls it.
// You see the resulting post_id and scheduled_for timestamp in the
// conversation. No SDK. No SDK upgrade when new tools ship.

Frequently asked questions