Live on Product Hunt Intraface launched today. Less steering, more vacations. See the launch
Start for free

MCP vs A2A: what's the difference?

MCP (Model Context Protocol) standardises how one AI agent reaches its tools, data and prompts. A2A (Agent2Agent) standardises how independent agents find each other, hand over tasks and return results. The A2A project describes them as complementary: MCP is vertical, deepening a single agent, and A2A is horizontal, linking agents across teams and vendors.

Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.

How it works

Your agent reaches its tools with MCP (down) and another agent with A2A (across); that agent uses MCP for its own tools.A2A: horizontal,agent to agentyouyour agentMCP client and A2A clientanother agentan A2A server elsewhereMCP serveryour toolsMCP serverits toolsMCP: vertical,agent to toolaskstools/call1. message opens a task2. status, artifactstools/call
  1. With MCP, an agent lists a server's tools, calls one with JSON arguments and gets a result. A tool is usually stateless: a lookup, a write, a query.

  2. With A2A, an agent first reads another agent's Agent Card, a JSON document giving its endpoint URL, capabilities, authentication needs and skills.

  3. It sends that agent a message. This opens a Task with its own id and lifecycle; the remote agent can stream updates over SSE or pause to ask for more input.

  4. The task ends in a terminal state (completed, canceled, rejected or failed) and hands back artifacts: the deliverables, as text, files or structured data.

  5. Inside the remote agent, MCP is probably doing the tool work. A2A leaves tools to MCP and says nothing about how an agent runs its own sub-agents.

Source: A2A Protocol: A2A and MCP

MCP versus A2A, side by side

Both protocols send JSON-RPC 2.0 over HTTP, both let one party discover what another offers, and both arrived as AI agents started talking to things other than people. That is where the resemblance ends.

MCP connects an agent to tools and resources. A tool has a name, a description and an input schema; the model calls it and gets a result. The A2A docs characterise tools as primitives with structured inputs and outputs that are often stateless, like a calculator or a weather lookup, and that is a fair picture of an MCP tool.

A2A connects an agent to other agents. Its unit of work is a Task: a stateful object with an id and a lifecycle, which may run for a long time, stream progress, stop in an input-required or auth-required state while it waits for the caller, and finish as completed, canceled, rejected or failed. Communication is a sequence of Messages, each from the user or the agent and made of Parts (text, a file or structured data), and the output is one or more Artifacts. Discovery is an Agent Card, a JSON document that names the agent's endpoint, its skills and the authentication it expects. To the caller the remote agent is opaque; its model, prompts and tools are not visible.

The A2A project's own summary is the clearest: MCP is vertical and deepens a single agent, A2A is horizontal and connects agents across an organisational boundary.

Example: the repair shop from the A2A docs

The A2A documentation uses a car repair shop. A customer talks to a shop manager agent over A2A and describes the problem across several turns. A mechanic agent uses MCP to call diagnostic scanners, pull up repair manuals and raise the lift. When a part is needed, the mechanic uses A2A again to ask a parts supplier agent, run by a different company, for a quote and an order. The conversational, task-shaped exchanges are A2A; the structured calls to equipment and databases are MCP.

When you need A2A at all

Most teams do not yet. A single agent with a good set of MCP servers covers a great deal, and sub-agents inside one product can be coordinated by the framework without a protocol between them. A2A earns its place when the other agent belongs to someone else: another team, another vendor, another trust boundary. Then you need a way to discover it, authenticate to it, hand it work that may take minutes, and get artifacts back without knowing how it is built. That is what an Agent Card and a Task lifecycle give you.

Common mistakes

  • Wrapping an agent as a tool and expecting multi-turn behaviour. An MCP tool call returns once. If the remote side needs clarification halfway through, you want a Task that can pause in input-required.
  • Using A2A between components you control. A protocol boundary costs you observability and speed. Inside one system, direct calls or MCP are simpler.
  • Reading "both use JSON-RPC" as "both do the same thing". The transport is shared; the objects on it (tools versus tasks) are not.

FAQ

Do MCP and A2A compete?
The A2A project says no, and the designs agree. MCP has no notion of a long-running task between peers; A2A has no notion of a tool list for a model. A system with several agents typically uses A2A between them and MCP inside each.
Who maintains A2A, and who maintains MCP?
A2A was started by Google and donated to the Linux Foundation, where a steering committee with members from AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow maintains it under the Apache 2.0 licence. MCP was created by Anthropic and is developed in the open at modelcontextprotocol.io.
Can an A2A agent be used as an MCP tool?
The A2A docs say an A2A server can expose some skills as MCP-compatible resources, and that this fits best when the skill behaves like a stateless tool call. Anything that needs a back-and-forth belongs on the A2A side.

Stop sending HTMLover Slack.

Create an account in a minute, connect your coding agent, and make your first board.