Communication protocols connecting AI models to tools and services.
As AI systems become more capable, they need standardized ways to communicate with tools, data sources, and other agents. AI protocols define these interfaces β from MCP (Model Context Protocol) which connects models to external tools via a universal standard, to A2A (Agent-to-Agent) for inter-agent communication, to the streaming protocols that deliver real-time responses.
These protocols are the "plumbing" of the AI ecosystem. Just as HTTP standardized web communication, protocols like MCP are standardizing how AI models interact with the world. Understanding them is crucial for building interoperable AI applications that can plug into the growing ecosystem of tools and services.
MCP (Model Context Protocol)
Anthropic's open standard for connecting AI to tools and data. Servers expose resources, tools, and prompts via a standard JSON-RPC protocol. One MCP server works with any MCP-compatible client (Claude, Cursor, etc.).
MCP Architecture
Client-server model: MCP hosts (AI apps) connect to MCP servers (tool providers). Servers declare capabilities. Transport layers: stdio (local), SSE/HTTP (remote). Stateful sessions with capability negotiation.
A2A (Agent-to-Agent Protocol)
Google's protocol for agent interoperability. Enables agents built on different frameworks to discover each other, negotiate capabilities, and collaborate on tasks. Complementary to MCP (tools) β A2A handles agent-agent communication.
OpenAI Function Calling
The first widely adopted tool-use protocol. Define functions with JSON Schema, model generates structured calls. Now an industry standard that Anthropic, Google, and others have adopted with variations.
Server-Sent Events (SSE)
One-way streaming protocol for delivering LLM tokens in real-time. HTTP-based, simple to implement, works through proxies. Used by OpenAI, Anthropic, and most API providers for streaming responses.
WebSocket AI Communication
Bidirectional streaming for real-time AI interactions. Used for voice AI (OpenAI Realtime API, Google Gemini Live), interactive agents, and scenarios requiring low-latency two-way communication.
OpenAI Assistants Protocol
Stateful conversation management with threads, runs, and built-in tools (code interpreter, file search). Higher-level than raw chat completions. Influenced how other providers design managed agent APIs.
Structured Output Standards
JSON Schema-based output enforcement across providers. OpenAI strict mode, Anthropic tool results, Google controlled generation. Converging toward a common standard for structured AI responses.
OAuth and Authentication
Emerging standards for secure tool access. MCP supports OAuth 2.0 for authenticated server connections. Key management, token rotation, and scope-based permissions for AI-to-service authentication.
The Protocol Landscape
MCP is for model-to-tool. A2A is for agent-to-agent. Function calling is for in-context tool use. SSE/WebSocket for streaming. The ecosystem is converging but not yet fully standardized β expect consolidation.
MCPModel Context Protocol β Anthropic's open standard for universal tool and data connectivity for AI models.
A2AAgent-to-Agent Protocol β Google's standard for enabling AI agents to discover and communicate with each other.
SSEServer-Sent Events β HTTP-based protocol for streaming LLM responses token by token in real time.
JSON-RPCRemote procedure call protocol using JSON β the transport mechanism underlying MCP communication.