Agent protocols: MCP and A2A
MCP connects an agent to tools and data; A2A connects an agent to other agents. Both are open specifications now hosted under the Linux Foundation, and both change often, so the versions below are dated.
This page assumes you know what an agent is. How agents are combined, with or without these protocols, is on Multi-agent systems.
MCP: connecting agents to tools
The Model Context Protocol (MCP) is an open standard for plugging tools and data into an AI app; it connects an agent to tools, not to other agents. The specification calls it "an open protocol that enables seamless integration between LLM applications and external data sources and tools" (MCP specification, checked 2026-10-28).
- Origin and governance. Anthropic open-sourced MCP in November 2024 (Anthropic announcement). In December 2025, "Anthropic is donating MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation" (MCP blog, checked 2026-10-28).
- Versions. The spec is versioned by date. "The current protocol version is 2026-07-28" (MCP versioning, checked 2026-10-28).
- Parts. An MCP server offers tools, resources (data) and prompts. An MCP client, inside an app such as a chat app or coding agent, connects to it. Messages use JSON-RPC 2.0.
- Transports. Two ways to connect (MCP
transports, checked 2026-10-28):
- stdio: the app starts the server as a local process and talks over standard input and output. Good for tools on your own machine.
- Streamable HTTP: "each message is an HTTP POST to a single MCP endpoint". This is how a remote MCP server runs, hosted like any web service. See the cloud for AI apps.
Example: a company runs an MCP server that exposes "search our tickets" and "create a ticket". Any MCP-capable assistant its staff use can then work with tickets, without a custom integration for each assistant.
One MCP server can itself call a model, so an "agent behind a tool" is possible. But MCP's design centres on tools for one agent, and its security questions (who may call which tool, with whose token) are covered on the security page.
A2A: agents talking to agents
Agent2Agent (A2A) is an open protocol for one agent to discover another agent, send it a task and follow the task's progress, across vendors and organizations. Its site calls it "an open standard for seamless communication and collaboration between AI agents", "originally developed by Google and donated to the Linux Foundation" (A2A, checked 2026-10-28).
- Agent Card. Each agent publishes "A JSON metadata document … describing its identity, capabilities, skills, service endpoint" so others can find out what it does and how to call it (A2A specification, checked 2026-10-28).
- Task. "The fundamental unit of work managed by A2A, identified by a unique ID." Tasks have a lifecycle (submitted, working, completed and so on), so long jobs can report progress.
- Version. The specification page lists 1.0.0 as the latest released version (checked 2026-10-28).
- Consolidation. IBM's Agent Communication Protocol (ACP) "is officially merging with the A2A under the Linux Foundation […] umbrella" (LF AI & Data, August 2025, checked 2026-10-28).
MCP and A2A are complementary: MCP is how an agent uses tools; A2A is how an agent asks another agent, which may use its own tools, to do a job.
A2A is younger than MCP in practice. Most teams today build multi-agent systems inside one program with a framework, and use A2A only when the other agent belongs to someone else.
Related pages
- Multi-agent systems
- What an AI agent is
- Security for AI apps: MCP risks
- The cloud, for AI apps: hosting a remote MCP server
See also
Glossary terms used on this page: A2A (Agent2Agent), Agent, Cloud, Coding agent, LLM, MCP (Model Context Protocol), Model, Prompt, Token, Tool.
Sources
All checked on 2026-10-28.
- Anthropic, Introducing the Model Context Protocol (November 2024)
- Model Context Protocol, Specification, Versioning, Transports
- MCP blog, MCP joins the Agentic AI Foundation (December 2025)
- A2A, Overview, Specification
- LF AI & Data, ACP joins forces with A2A (August 2025)
First version written with AI assistance and checked against the sources on 2026-10-28. Found a mistake? Fix it or open an issue. Text licensed under CC BY 4.0.