A2A and MCP are not competitors. The Model Context Protocol (MCP) connects an AI agent to tools and data: databases, SaaS apps, files and APIs. The Agent2Agent protocol (A2A) connects agents to other agents, so one agent can discover another, hand it a task and get the result back without either exposing its internal tools or memory. A common way to put it: MCP works vertically, from an agent down to its tools, and A2A works horizontally, between agents.
Most serious agent systems will use both. Use MCP to give each agent its capabilities, and A2A when agents built by different teams, vendors or companies need to work together. If you are new to MCP, start with our explainer on what MCP is.
| MCP | A2A | |
| Connects | An agent to tools, data and prompts | An agent to other agents |
| Direction | Vertical: agent to tool | Horizontal: agent to agent |
| Created by | Anthropic, November 2024 | Google, April 2025 |
| Governance | Agentic AI Foundation, Linux Foundation | Linux Foundation, now within the Agentic AI Foundation |
| Discovery | Clients ask a server what tools it offers; public servers listed in the MCP Registry | Agents publish an agent card at a well-known address |
| Unit of work | A tool call with arguments and a result | A task with status, messages and artifacts, which can run for a long time |
| Visibility | The agent sees the tool’s interface | Agents stay opaque to each other |
| Latest milestone | 2026-07-28 specification | Version 1.0 in March 2026, with signed agent cards |
| Typical use | Let an agent query a CRM or open a ticket | Let a travel agent ask a payments agent to book and pay |
MCP standardizes how AI applications reach tools and data. An MCP server exposes tools, resources and prompts, and any MCP-compatible client, such as Claude, ChatGPT, Cursor, Codex or Copilot, can connect to it. The agent decides when to call a tool, and the server does the work, usually by calling an API. Our comparison of MCP vs function calling vs APIs explains how it fits with older approaches.
A2A standardizes how agents collaborate. Each agent publishes an agent card, a JSON document describing who it is, what skills it offers, where to reach it and how to authenticate. Another agent reads the card, sends a task and follows its progress through statuses such as submitted, working and completed, receiving messages and output files along the way. Agents never need to share their prompts, memory or tools, which matters when they belong to different companies.
Google announced A2A in April 2025 and donated it to the Linux Foundation that June. Version 1.0, released in March 2026, added cryptographically signed agent cards so a receiving agent can verify the card really came from the domain owner, along with multi-tenancy and several protocol bindings. More than 150 organizations support the project.
Picture a company’s procurement agent. It uses MCP to read the inventory database and the approved supplier list. When stock runs low, it uses A2A to find a supplier’s sales agent, send a purchase request and track the order. The supplier’s agent, in turn, uses MCP to check its own warehouse and pricing systems. Payment can ride on A2A too: Google’s Agent Payments Protocol is an A2A extension, and its x402 extension lets agents settle in stablecoins. Our guide on what x402 is covers that piece.
MCP security centers on what an agent can reach: which servers it trusts, what each tool may do and how user credentials are handled. A2A security centers on who an agent is talking to: verifying agent cards, authenticating the remote agent and deciding which tasks it may accept or delegate. Both are exposed to prompt injection, since content from a tool or another agent can carry hidden instructions. Our AI agent security checklist covers controls for both.
MCP servers are listed in the official MCP Registry and directories compared in our list of MCP directories and registries. A2A agents are found through agent cards at well-known URLs and through registries, from cloud catalogs at Google, Microsoft and AWS to public directories, covered in our list of AI agent registries and directories.
No. They solve different problems and are designed to be used together. Both now sit under the Linux Foundation’s agentic AI umbrella.
Usually not. A single agent with the right MCP servers covers most internal use cases. A2A becomes valuable when several agents, especially from different owners, need to collaborate.
Google created it, and the project lists support from more than 150 organizations, with founding partners including AWS, Cisco, Microsoft, Salesforce, SAP and ServiceNow.
Yes, and that is the common pattern: it uses MCP for its own tools and A2A to talk to other agents.
MCP gives an agent its hands, the tools and data it can use. A2A gives it colleagues, other agents it can find, trust and delegate to. Start with MCP to make each agent useful, then add A2A when your agents need to work with agents you do not own. With both protocols now under neutral governance, building on them is a safe long-term bet.