MCP vs A2A vs NLWeb
MCP vs A2A vs NLWeb

Agentic Protocols: MCP · A2A · NLWeb

Agentic AI Infrastructure

MCP · A2A · NLWeb
The Three Agentic Protocols

A visual guide to the foundational protocols enabling AI agents to use tools, collaborate with each other, and interact with the web.

Model Context Protocol Agent-to-Agent Natural Language Web

What Each Protocol Does

🔌

MCP

Model Context Protocol · by Anthropic (2024)

Standardizes how LLMs connect to external tools, APIs, databases and resources. Like a USB-C port — one universal interface for any tool or data source.

Tool Use Resources Prompts Client-Server
🤝

A2A

Agent-to-Agent Protocol · by Google (2025)

Enables autonomous AI agents to discover, communicate and delegate tasks to one another across different frameworks and vendors. Agents as a network.

Multi-Agent Delegation Discovery Interop
🌐

NLWeb

Natural Language Web · by Microsoft (2025)

Adds a conversational interface layer to any website. Sites expose structured data via schema.org so AI agents can query and navigate them in natural language.

Web Agents schema.org RSS/HTML Q&A API

How Each Protocol Flows

🔌 MCP — Model Context Protocol Flow

👤 User / App 🧠 MCP Host Claude / GPT / Gemini + MCP Client ⚙️ MCP Server Exposes tools, resources, prompts 🗃 Database 🔎 Search API 📁 File System query tool_call result returned to LLM request response

🤝 A2A — Agent-to-Agent Protocol Flow

👤 User Task Request 🎯 Orchestrator Agent A Plans & delegates sub-tasks 📋 Agent Cards Registry / Discovery /.well-known/agent.json 🔍 Research Agent B ✍️ Writer Agent C 📊 Analyst Agent D task discover delegate sub-tasks results

🌐 NLWeb — Natural Language Web Flow

👤 User or AI Agent 💬 NL Query “What’s the best laptop under $1k?” 🌐 NLWeb Layer site.com/nlweb Parses query + LLM schema.org data 📄 schema.org 🔗 RSS / Atom 🗂 Site Content ask HTTP POST 📦 JSON-LD Structured answer structured response back to agent / user

Protocol Examples

MCP · Code Assistant + Git

User asks Claude to “Fix the failing tests in my repo and push a commit”

1
Claude receives request (MCP Host)
2
Calls filesystem/read tool → reads test files
3
Calls shell/run → runs pytest, gets errors
4
Calls filesystem/write → applies fixes
5
Calls git/commit_push → done ✓
MCP · Customer Support Bot

User: “What’s the status of my order #8821?”

1
LLM detects intent: order lookup
2
MCP Server receives get_order(id: "8821")
3
Queries internal DB → returns JSON order object
4
LLM formats into natural language reply
5
“Your order is out for delivery, arriving today” ✓
A2A · Research & Report Pipeline

User: “Write me a market analysis report on EVs”

1
Orchestrator Agent breaks task into subtasks
2
Discovers Research Agent via agent.json card
3
Delegates “gather EV data” → Research Agent responds with findings
4
Delegates “write report” to Writer Agent with research context
5
Orchestrator compiles + delivers final report ✓
A2A · Travel Booking System

User: “Plan and book a 5-day Tokyo trip for 2”

1
Travel Orchestrator Agent receives request
2
Flight Agent searches & books flights (parallel)
3
Hotel Agent books accommodation (parallel)
4
Itinerary Agent creates day-by-day plan
5
Orchestrator returns complete trip package ✓
NLWeb · E-commerce Assistant

Agent: “Find a waterproof hiking boot under $150 size 11”

1
Agent POSTs NL query to boots.com/nlweb
2
NLWeb layer parses query, matches schema.org Product type
3
Filters by price, waterproof attribute, size=11
4
Returns JSON-LD structured list of matching products
5
Agent presents options to user or auto-purchases ✓
NLWeb · News Aggregation Agent

Agent: “Summarize today’s AI news from major tech blogs”

1
Agent queries multiple sites via NLWeb endpoints
2
Each site’s NLWeb returns schema.org Article objects
3
Agent aggregates structured data (no HTML scraping needed)
4
Deduplicates & ranks by relevance
5
Delivers clean summary with citations ✓

Protocol Comparison

Attribute MCP A2A NLWeb
Created by Anthropic (2024) Google (2025) Microsoft (2025)
Primary purpose LLM ↔ Tools/Data Agent ↔ Agent Agent ↔ Web
Transport stdio, HTTP SSE HTTP/REST, SSE HTTP POST/GET
Data format JSON-RPC 2.0 JSON (Task objects) JSON-LD, schema.org
Discovery Pre-configured servers Agent Cards (.well-known) Site’s /nlweb endpoint
Best for Tool integration, DB access, file ops Complex multi-agent workflows Web browsing agents, e-commerce
Auth OAuth 2.0, API Keys OAuth 2.0, JWT Standard HTTP auth
Analogy USB-C for AI tools HTTP for agents RSS/APIs for natural language

All Three Working Together

Combined Agentic Stack — “Plan my weekend trip and blog about it”

👤 User “Plan trip & blog” A2A Orchestrator Breaks task, delegates agents NLWeb Travel Agent Queries hotel/flight sites MCP Writer Agent Uses filesystem + CMS tools Budget Agent A2A sub-agent hotels.com/nlweb flights.com/nlweb 📁 filesystem MCP 📝 WordPress MCP ✅ Done! Trip booked + blog published

MCP by Anthropic · A2A by Google · NLWeb by Microsoft · Illustrated 2025

Leave a Reply

Your email address will not be published. Required fields are marked *