MCP vs A2A
Different layers of the same stack. MCP connects an agent downward to its tools and data — one agent, many capabilities. A2A (Agent2Agent) connects agents sideways to each other — so an agent built by one vendor can hand a task to an agent built by another. They were designed to be used together, and serious multi-agent systems use both.
Side by side
| MCP | A2A | |
|---|---|---|
| Connects | An agent to tools and data sources | Agents to other agents |
| Origin | Anthropic, November 2024 | Google, April 2025 |
| Governance today | Agentic AI Foundation (Linux Foundation), since December 2025 | Linux Foundation, since June 2025; v1.0 shipped in 2026 |
| Discovery | tools/list — typed tool schemas | Agent Cards — published descriptions of an agent's skills |
| Unit of work | A tool call: request in, result out, seconds | A task: delegated, tracked, possibly long-running |
| Typical question | "Can my assistant query our database?" | "Can our support agent hand this case to the billing agent?" |
When to use which
One assistant that needs capabilities? MCP. Connecting Claude or Cursor to GitHub, Notion or your database is agent-to-tool — A2A adds nothing here. This covers the overwhelming majority of real usage today.
Multiple autonomous agents that must cooperate? A2A. When a procurement agent negotiates with vendor agents, or agents from different platforms hand off work, they need a common language for capabilities, task state and results between peers.
Building an agent platform? Both, at different boundaries: A2A at the edges where your agents meet the world's agents, MCP inside, where each agent touches actual systems.
Common questions
Are MCP and A2A competitors?
No — Google positioned A2A as complementary to MCP from its very first announcement, and both now live under Linux Foundation umbrellas. MCP is the agent-to-tool layer; A2A is the agent-to-agent layer. A common architecture uses A2A to delegate a task between agents, each of which uses MCP to do the actual work.
Who governs each protocol now?
MCP was created by Anthropic (November 2024) and donated to the Agentic AI Foundation — a Linux Foundation initiative co-founded with OpenAI and Block — in December 2025. A2A was announced by Google in April 2025 and transferred to the Linux Foundation in June 2025, where AWS, Cisco, Microsoft, Salesforce, SAP and ServiceNow are among the members.
Do I need A2A?
Only if independent agents — usually from different vendors or teams — need to discover each other and collaborate. A single assistant with many capabilities is an MCP problem, not an A2A one. Most people asking this question need MCP first.
Can an A2A agent use MCP servers?
Yes, and that's the intended shape: A2A handles the negotiation between agents — who does what, task status, results — while each agent reaches its own databases, APIs and tools through MCP.
More pairings: the compare hub · protocol basics: What is MCP?