lockstep
4 min readNaman Jain

MCP vs A2A vs ACP: agent protocols explained

MCP vs A2A vs ACP compared: MCP connects agents to tools and data, A2A connects agents to other agents, and ACP connects editors to coding agents.

An architecture review has a slide titled "Our agent protocol strategy" with three acronyms on it: MCP, A2A, ACP. Someone asks which one to pick. The answer is that they aren't alternatives. Each connects a different pair of things.

This guide explains them from their own sources: the MCP architecture docs, the A2A project, and the Agent Client Protocol, with Kiro's ACP docs, checked on 16 November 2026.

The short answer

MCPA2AACP
Full nameModel Context ProtocolAgent2Agent ProtocolAgent Client Protocol
ConnectsAn AI application to tools and dataAn agent to other agentsA code editor to a coding agent
Question it answersWhat can the agent use?Who can the agent work with?How does a user drive the agent?
ExampleClaude Code calling a Sentry or database serverA travel agent asking a booking agent to reserve flightsZed or a JetBrains IDE running Kiro as its agent

MCP: agent to tools

MCP uses a client-server architecture. An MCP host, an AI application such as Claude Code, Claude Desktop, or VS Code, creates one MCP client per server it connects to, and each server provides context: tools, resources, and prompts. Local servers typically use the stdio transport and serve one client; remote servers use Streamable HTTP and serve many.

Use it to give an agent access to systems: your database, issue tracker, docs, or internal APIs. See Claude Code MCP setup and MCP vs RAG.

A2A: agent to agent

The A2A project describes an open protocol for communication between opaque agentic applications: agents built on different frameworks, by different companies, on different servers, working together as agents rather than as tools. Per its README, agents can:

  • Discover each other's capabilities
  • Negotiate how to interact (text, forms, media)
  • Collaborate on long-running tasks
  • Do so without exposing internal state, memory, or tools

Technically: JSON-RPC 2.0 over HTTP(S), Agent Cards for discovery, and support for request and response, streaming over server-sent events, and asynchronous push notifications. The project says A2A complements MCP.

Use it when separately built agents need to delegate work to each other across organizational or system boundaries.

ACP: editor to coding agent

The Agent Client Protocol standardizes communication between code editors and coding agents. Its current stable protocol version is 1, with versions negotiated at initialization.

Kiro shows it in practice: its IDE, CLI, web, and mobile apps are all ACP clients of one agent harness, and compatible editors such as JetBrains IDEs and Zed can use Kiro the same way. Kiro's docs summarize the split: ACP connects a client to the agent and answers how a user interacts with and controls it; MCP connects the agent to tools and data.

Use it if you're building a coding agent that editors should drive, or an editor that should support many agents.

How they fit together

One coding agent can use all three:

Editor  --ACP-->  Coding agent  --MCP-->  Tools and data
                       |
                       +--A2A-->  Other agents (e.g. a security review agent)
  • ACP on the front: the user drives the agent from their editor.
  • MCP on the back: the agent reaches tools and data.
  • A2A on the side: the agent delegates to other agents it doesn't control.

Which matters for your team

  • Using coding agents: MCP. It's how you give agents your systems.
  • Building multi-agent systems across vendors: A2A.
  • Building an editor or a coding agent: ACP.

What none of them carries

The protocols move requests and results. None of them defines what your team has decided, such as which API version agents must use or which vendor was rejected. That has to come from a source agents can query.

Where Lockstep fits

Lockstep is that source, exposed over MCP. It captures decisions from PRDs, Slack, and pull requests, routes each to its owner for approval, and serves them through an MCP server any MCP client can call. Claude Code sessions are also briefed automatically on the decisions that apply.

See how it works.

Sources checked 16 November 2026

For MCP security, see MCP security. For the agent terms behind these protocols, see agentic AI vs AI agents vs generative AI. To build an MCP server, see how to build an MCP server. For a hands-on A2A walkthrough, see A2A protocol example.

Frequently asked questions

What is the difference between MCP and A2A?
MCP connects an AI application to tools, data, and prompts that servers provide. A2A connects agents to other agents, so agents built by different companies on different frameworks can discover each other's capabilities and collaborate on tasks, without exposing their internal memory or tools. The A2A project describes it as complementing MCP.
What is the Agent Client Protocol (ACP)?
A protocol that standardizes communication between code editors and coding agents. Kiro, for example, uses it so its IDE, CLI, web, and mobile apps, and editors such as JetBrains IDEs and Zed, can all drive the same agent.
What is the difference between ACP and MCP?
They sit on opposite sides of the agent. Kiro's docs put it this way: ACP connects an editor or other client to the agent and answers how a user interacts with and controls it; MCP connects the agent to external tools and data and answers which systems it can use.
How does A2A work?
Per its README, A2A uses JSON-RPC 2.0 over HTTP(S), Agent Cards that describe an agent's capabilities and connection details for discovery, and supports request and response, streaming with server-sent events, and asynchronous push notifications, exchanging text, files, and structured JSON.
Do I need all three protocols?
Usually not. Most teams building on coding agents only need MCP, to give agents tools and data. A2A matters when separate agents must work together across systems. ACP matters if you're building an editor integration or a coding agent that editors should drive.

Keep reading