Tool Calling VS MCP in AI Agents | Model Context Protocol Explained

Technological Concepts Explained

LLMs as a “black box”

Large language models (LLMs) primarily take input tokens and predict output tokens. By themselves, they can’t reliably:

• Search the internet in real time
• Query or mutate a user’s databases
• Perform external actions (e.g., make phone calls)

Tool calling / function calling (traditional agent approach)

In the traditional approach:

• Developers define custom functions/tools (e.g., GetWeather(city)), typically backed by APIs.
• The LLM can decide to call these tools when answering, enabling capabilities like:
  • Internet-backed weather retrieval
  • Database read/write
  • Other external integrations

The problem with tool calling: no standard tool definition format

A tool schema/function definition that works for one provider may not work for another (e.g., OpenAI vs. Gemini or Claude). As a result, tools are:

• Not easily portable/shareable across different LLM ecosystems

Model Context Protocol (MCP)

MCP aims to standardize how tools are described and exposed:

• Tools implemented as MCP servers/tools can be used by different LLMs and clients (OpenAI, Gemini, etc.) because they follow the same protocol.
• This creates a more “plug-and-play” ecosystem—one tool definition can be reused across clients.

---

Product / Features Highlighted (Sponsored: “Storm MCP”)

Storm MCP provides a gateway

Storm MCP sets up a gateway that exposes a collection of MCP-connected tools behind a single URL.

• Multiple tools/services can be grouped behind that gateway (e.g., Cloudflare DNS/analytics, Linear, search providers, etc.).

Authentication layer

Storm MCP adds an authentication step beyond the gateway.

Example flow shown:

1. Connect the Cloudflare DNS MCP server to a user’s Cloudflare account (permission prompt).
2. Confirm “authentication successful.”

Tool aggregation into clients (Cursor / IDE integration)

Storm MCP includes an action similar to “Add to Cursor”:

• It generates an API key / config to enable the toolset in Cursor.
• After enabling it, the LLM can automatically use tools such as:
  • AccountList / Set Active Account / Zone List
  • DNS Report (to fetch DNS entries/records)
  • Linear operations like get issues, list issue labels, update issue, etc.

Demonstrated autonomous tool use

In the demo:

• The speaker asks, “what domains I own,” and the system automatically uses the relevant tools:
  • lists accounts/zones
  • then retrieves DNS information
• After permissions are connected, it fetches DNS report details to confirm tool-driven access works.

---

Guide / Tutorial Workflow Shown in the Video

1. Connect an MCP server to an external service (e.g., Cloudflare) with user approval.
2. Deploy a gateway in Storm MCP and select which tools/apps to include.
3. Add the gateway to an LLM client/IDE (Cursor) using API authentication.
4. Verify tool availability inside the chat (e.g., confirm tools appear—7 tools for Cloudflare in the demo).
5. Run tasks, such as:
   • Domain enumeration → DNS report/records
   • Connect more apps (e.g., Linear) → issue/PR retrieval and updates

---

Key Analysis / “Difference” Takeaway

• Tool calling: Works, but interoperability is limited because tools are tied to each provider’s specific function/schema format.
• MCP: Provides a standard protocol, making tools portable across LLMs/clients and reducing duplication/rework while enabling broader reuse.

---

Main Speakers / Sources

• Speaker: Piyush (as referenced in subtitles during the demo; mentions like “Piyush Kar…” / “Piyush Kar dot pro”)
• Sponsored product / source: Storm MCP
• Examples discussed: OpenAI tool calling/function calling, Gemini, Claude, plus external services like Cloudflare and Linear