If you've been following some of the innovations in AI, you've probably heard about the Model Context Protocol (MCP) and Agent Client Protocol (ACP). While they sound similar, they solve different problems in the agentic stack.

The agentic stack has three layers: the LLM (the brains that process text), the Agent (the orchestrator that executes code and uses tools), and the Client (the application humans interact with, like VS Code or a web app).

It is important to understand that an LLM can only produce text—it can't read files or call APIs directly. Consequently, given a prompt such as “What is the weather in Seattle”, the LLM tells the agent to “call the weather tool” with location = Seattle. The agent executes the command requested and returns the results back to the LLM which uses the results to formulate a nice response: “It's sunny in Seattle” (yeah right!). 

MCP (Model Context Protocol) standardizes this agent-to-tool communication, ensuring tools work consistently across different agents and LLMs.

On the other end, clients such as text editors need to communicate with agents. The Agent Client Protocol (ACP) standardizes this client-to-agent communication, so any client can work with any agent.

So the ACP standardizes the integration at the top of the stack while the MCP standardizes the integration at the bottom of the stack. Both protocols allow standardized services to be exposed using a common contract and all of it runs over HTTP (services exposed over a network) or stdio (for local processes).

The more things change, the more they stay the same.