MCP vs A2A: What Each Protocol Does and When to Use Both
Updated 2026-11-18
MCP (Model Context Protocol) connects an AI agent to tools, APIs and data. A2A (Agent2Agent) connects one agent to other agents so they can discover each other and work on tasks together. They solve different problems and are designed to be used together: MCP for what an agent can use, A2A for who an agent can work with.
What MCP is for
MCP is an open-source standard for connecting AI applications to external systems. A server exposes tools (functions the model can call), resources (data) and prompts (templates). A client in the host application keeps a connection to each server. A typical call looks like: the model decides to run search_issues, the client sends the call to the server, and the server returns a result.
The A2A documentation describes the relationship in two lines: "MCP is vertical. It deepens a single agent." MCP tools are used by an agent, and they do not reason or hold a conversation of their own. If you are new to it, read what an MCP server is.
What A2A is for
A2A is an open standard for AI agents to talk to each other, whatever framework or vendor built them. It is built on existing web standards: HTTP, JSON-RPC and Server-Sent Events.
Its main pieces, as the A2A docs describe them:
- Agent Card: a JSON description of an agent's endpoint, capabilities, security schemes and skills. Clients fetch it from a well-known path, for example
/.well-known/agent-card.json. - Task: created by the server when it processes a message. Tasks move through states such as submitted, working and completed.
- Message: how a client sends a request, either as a single request and response or as a stream of updates.
- Artifact: the output a task produces.
A2A also supports long-running work through streaming and push notifications. Agents collaborate without exposing their internal logic, memory or proprietary tools. In the docs' words: "A2A is horizontal. It connects agents across that boundary."
Side-by-side comparison
| MCP | A2A | |
|---|---|---|
| Connects | An agent to tools, resources and data | An agent to other agents |
| Direction | Vertical: deepens one agent | Horizontal: connects agents across boundaries |
| Typical interaction | Structured calls, often one request and one result | Multi-turn collaboration on tasks that can run for a long time |
| What the other side is | A program that exposes functions | An agent that reasons and keeps its internals private |
| Discovery | Registry listings, client config, server tools/list | Agent Card at /.well-known/agent-card.json, registries or direct config |
| Transports | stdio (local) and Streamable HTTP (remote) | HTTP, JSON-RPC and SSE |
| Example | Query a database, create a payment link | Ask a billing agent to handle a refund dispute |
An example of using both
The A2A documentation uses an auto repair shop to show how the two fit together. Say the shop is staffed by AI agents:
- A customer tells the shop manager agent about a rattling noise. They hold a diagnostic conversation. That is A2A: agents talking to agents.
- The mechanic agent uses a diagnostic scanner, a repair manual database and a platform lift, each through structured calls. That is MCP: an agent using tools.
- The mechanic asks a parts supplier agent whether a part is in stock, and that can lead to an order. That is A2A again, across an organization boundary.
A software version: a support agent at your company uses MCP to read your ticket system and knowledge base. When a request involves a payment dispute, it delegates to a payments agent run by another team or vendor over A2A. That agent uses its own MCP servers to look up transactions. You see only the result.
When to use which
| You want to... | Use |
|---|---|
| Let Claude, ChatGPT, Cursor or VS Code call your product's functions | MCP server |
| Expose a database, file store or internal API to an agent | MCP server |
| Let other organizations' agents delegate tasks to your agent | A2A endpoint with an Agent Card |
| Coordinate several specialist agents built on different frameworks | A2A between them, MCP inside each |
| Do both: sell a capability and also run an agent that uses it | MCP for the tools, A2A for the agent |
For most teams, MCP is the starting point because the clients that exist today, such as Claude Code, Cursor and ChatGPT, support MCP servers. A2A matters when your product is itself an agent that others need to call.
Discovery for both
Both protocols need to be found. For MCP, publish to the official registry. For A2A, publish an Agent Card at the well-known path. The A2A docs say a curated registry can also hold cards, but the current specification does not prescribe a standard API for curated registries. Our guide to getting your AI agent discovered covers both. Agents Index currently lists MCP servers from the official registry. Browse them in the official registry listing.
Sources
- A2A and MCP - a2a-protocol.org
- What is A2A? - a2a-protocol.org
- Agent Discovery - a2a-protocol.org
- What is MCP? - modelcontextprotocol.io
- MCP architecture overview - modelcontextprotocol.io
FAQ
What is the difference between MCP and A2A?
MCP connects one agent to tools, APIs and data. A2A connects agents to each other so they can discover, delegate and collaborate. The A2A docs call MCP vertical and A2A horizontal.
Does A2A replace MCP?
No. The A2A documentation calls the two complementary and recommends using them together: MCP gives each agent depth, and A2A gives a system of agents reach.
How does an A2A agent get discovered?
Through its Agent Card, a JSON file usually served at /.well-known/agent-card.json. Clients can also use a curated registry or hardcoded configuration.
Which one should I build first?
If you want AI apps to call your product's functions, build an MCP server. If you run an agent that other agents should delegate work to, build an A2A endpoint with an Agent Card.