---
title: "Discussion on \"Interfacing AI: Deciding Between MCP and CLI in Agentic Workflows\" | Hashnode"
description: "Discussion on \"Interfacing AI: Deciding Between MCP and CLI in Agentic Workflows\". Interfacing AI: Deciding Between MCP and CLI in Agentic Workflows\nIntroduction: The \"Brain\" needs a \"Hand\"\nAs we transition from simple \"Chatbots\" to autonomous \"Agents,\" the most critical design deci"
url: "https://hashnode.com/posts/interfacing-ai-mcp-vs-cli/6a002c4b52e99ac3b4c31e3f"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Interfacing AI: Deciding Between MCP and CLI in Agentic Workflows](https://blog.mukundmurali.in/interfacing-ai-mcp-vs-cli?utm_source=hashnode&utm_medium=feed)

Interfacing AI: Deciding Between MCP and CLI in Agentic Workflows Introduction: The "Brain" needs a "Hand" As we transition from simple "Chatbots" to autonomous "Agents," the most critical design deci

May 10 · 5 min read · blog.mukundmurali.in  
Read the article

-   [#ai-agents](https://hashnode.com/tag/ai-agents)
-   [#cli](https://hashnode.com/tag/cli)
-   [#llm](https://hashnode.com/tag/llm)
-   [#mcp](https://hashnode.com/tag/mcp)

## Written by

[Mukund Murali](https://hashnode.com/@mukundmurali)

@mukundmurali · Bengaluru, India

  
[View profile](https://hashnode.com/@mukundmurali)

Technical Product Manager @ LSEG | Exploring the intersection of LLMs, Agentic Workflows, and Cloud Infrastructure. I write about building smarter systems, the Model Context Protocol (MCP), and the art of compounding knowledge via a digital second brain.

## Responses 1

Markdown supported · ⌘↵ to post

[Deniafe](https://hashnode.com/@deniafe)

[Oct 10](https://hashnode.com/posts/interfacing-ai-mcp-vs-cli/6a002c4b52e99ac3b4c31e3f/comment/6aca33ff5d24316c219a009d)

The Token Tax point is incredibly relevant right now. I see this happening constantly in web automation, people pass massive, entire DOM structures to vision models or agents just to click a button, and the API costs explode. I actually ended up building a visual scraper (Selonate) just to avoid this, relying on traditional code for navigation to save tokens, and only using AI locally when a selector fails. Great breakdown on the context engineering overhead!
