Protocol
MCP vs A2A
MCP and A2A are both open protocols stewarded by the Linux Foundation, both built on JSON-RPC 2.0, and both released in 2024 and 2025. They sit at different layers of an agentic system and they often run side by side.
The distinction
The Model Context Protocol (MCP) connects one agent to its tools and data sources. It is the private tool-access layer for an individual agent. An MCP server exposes Resources, Tools, and Prompts; the agent client invokes them over JSON-RPC 2.0.
The Agent2Agent Protocol (A2A) connects multiple agents to each other. It is the public routing and delegation layer between agents. One agent calls another agent to perform a sub-task, instead of calling a tool directly.
How A2A got here
Google released A2A on April 9, 2025 at Google Cloud Next. The protocol was donated to the Linux Foundation on June 23, 2025; the LF Agent2Agent Protocol Project launched the same week, sitting alongside MCP inside the Linux Foundation's Agentic AI Foundation. A2A reached version 1.0 with 150+ adopting organizations by mid-2026. Backers include AWS, Microsoft, Salesforce, Cisco, SAP, and ServiceNow.
A2A's primitives
- Agent Cards. Cryptographically signed capability advertisements. An agent publishes a card describing what it can do; other agents read it to confirm identity and decide whether to delegate.
- Tasks. The unit of delegated work. One agent submits a Task to another over JSON-RPC 2.0; the receiving agent runs it, streams progress, returns the result.
- Transport. HTTP plus Server-Sent Events (SSE) plus JSON-RPC 2.0, the same wire pattern MCP uses for remote servers.
Shared design choices
| MCP | A2A | |
|---|---|---|
| Layer | Agent to its own tools and data | Agent to other agents |
| Released | Anthropic, Nov 2024 | Google, Apr 9, 2025 |
| Governance | Linux Foundation AAIF (Dec 2025) | Linux Foundation AAIF (Jun 23, 2025) |
| Wire format | JSON-RPC 2.0 over stdio, SSE, or HTTP | JSON-RPC 2.0 over HTTP plus SSE |
| Primitives | Resources, Tools, Prompts | Agent Cards, Tasks, transport |
| Trust signal | Client-side approval of tool calls | Signed Agent Cards advertise identity |
How they compose
In a multi-agent system, the two protocols run together. MCP handles each agent's tool access locally. A2A handles delegation between agents, where one agent hands a Task to another agent that owns a different toolset or a different model. A planning agent delegates a research Task to a research agent over A2A; the research agent invokes its own MCP-connected search and database servers to satisfy the Task; the result returns over A2A.
When to use which
- Use MCP when you are building one agent that needs many tools and data sources. Each integration is an MCP server; the agent invokes tools at inference time.
- Use A2A when you are orchestrating multiple agents that hand work to each other. Each agent publishes an Agent Card; delegation flows over Tasks.
- Use both when production calls for specialist agents with private toolsets coordinating on shared workflows. MCP wires each agent to its tools; A2A wires the agents to each other.