MCP vs A2A: what each protocol is for

What is MCP, and what is A2A?

MCP, the Model Context Protocol, is the tool layer. In the words of the A2A project's own documentation, it defines how an AI agent uses individual tools and resources, such as a database or an API. Anthropic introduced it on November 25, 2024.

A2A, the Agent2Agent Protocol, is the collaboration layer. It lets different agents collaborate to reach a common goal. Google launched it on April 9, 2025 with support from more than 50 technology partners, and on June 23, 2025 the Linux Foundation announced the Agent2Agent project to host it.

The A2A documentation draws the line in one sentence: “A2A connects the agents to each other; MCP connects each agent to its own tools.”

How do MCP and A2A compare?

Sources for the rows: the MCP specification and its transports page, Anthropic's donation announcement, Google's A2A announcement (client and remote agents, tasks, artifacts, Agent Cards, and the standards it builds on), and the Linux Foundation press release.

What does it look like when both are used?

The A2A documentation uses an auto repair shop staffed by AI agent mechanics. A customer tells the shop that the car is making a rattling noise. In that picture, the conversation between the customer and a mechanic, or between two mechanics, is agent-to-agent work: open-ended, back and forth, with a goal. That is A2A.

Each mechanic also uses special-purpose tools to diagnose and repair the problem. Using a tool is a discrete request with a structured answer. That is MCP.

Translate that to a software company. A customer's assistant asking your support agent to resolve a billing dispute would be A2A. Your support agent looking up the invoice and issuing the refund through your billing system is MCP.

Which one do you need first?

For most companies the answer is MCP, for a practical reason: tools come before delegation.

• You want your product usable from Claude, ChatGPT, Cursor or VS Code. Build an MCP server. Those clients connect to tools over MCP; Anthropic's December 2025 post lists them among the products that adopted it.
• You are building one agent that needs to reach internal systems. Give it MCP servers for those systems.
• You run several independent agents, possibly from different vendors, that must hand work to each other. That is where A2A fits, on top of the tools each agent already has.

What do people get wrong about MCP and A2A?

• Treating them as rivals. Google introduced A2A as a complement to MCP, and the A2A documentation's advice is “Use both together.”
• Wrapping an agent as a tool to avoid a decision. It can work for a simple case, but a tool call is one request and one answer. Long-running work with progress and hand-offs is what the A2A task lifecycle is for.
• Waiting for the standards to settle. Both protocols are now governed under the Linux Foundation. A small MCP server with a few well-described tools is useful today and does not lock you out of A2A later.

Next: why MCP exists, or how to build a server for your company.

Sources

Every figure and dated fact above comes from one of these pages. Follow the link to check it.

1. A2A and MCP. A2A Protocol documentation.
2. Announcing the Agent2Agent Protocol (A2A). Google for Developers, April 9, 2025.
3. Linux Foundation Launches the Agent2Agent Protocol Project. The Linux Foundation, June 23, 2025.
4. Introducing the Model Context Protocol. Anthropic, November 25, 2024.
5. Donating the Model Context Protocol and establishing the Agentic AI Foundation. Anthropic, December 9, 2025.
6. Model Context Protocol specification, version 2025-06-18. modelcontextprotocol.io.
7. MCP specification: Transports. modelcontextprotocol.io.