Agent Client Protocol

Programmers usually write their code in a special text program, the editor. For a few years now, an AI assistant has often been sitting inside these editors, suggesting code, hunting for bugs, or restructuring entire files. The problem: there are many different editors and many different assistants. Building each combination individually would be an enormous amount of work. The Agent Client Protocol is a fixed agreement about which messages the two sides exchange. It was published by the company behind the Zed editor and is freely usable.

Why editors and AI assistants need a common language

Without such an agreement, an effort problem arises. With five editors and five assistants, one would theoretically have to program 25 connections. With a protocol, that drops to ten: each side only has to master the one shared language. After that, everything fits together automatically.

For users, this means real freedom of choice. You can switch the assistant without giving up the editor you’re used to. Conversely, you can switch the editor and take your assistant with you. Until now, many such assistants were tightly bound to a single product.

Economically, this is a point of contention. Whoever offers a closed connection binds customers more strongly to themselves. Open protocols weaken this lock-in and increase competition. That is exactly why the industry frequently debates who supports which standards.

Which messages travel back and forth between editor and agent

Technically, two programs run at the same time. The editor is the client, meaning the party placing the request. The assistant runs as its own program alongside it and is called an agent, because it independently carries out several work steps in sequence. Both send each other short messages in a fixed format, similar to filled-out forms.

A typical sequence looks like this. The editor reports which file is open and what the user wants. The agent responds with requests, such as: read this file to me, or: change line 40. The editor carries out these requests and reports the result back. For risky steps, such as deleting files, the editor first asks the user for confirmation.

The division of roles is important. The agent thinks, but it does not access the computer directly. All changes go through the editor, which acts as the controlling authority. The protocol is built on the widespread JSON-RPC format, with which programs call functions on the other side via text messages. ACP should not be confused with the Model Context Protocol, or MCP for short: that one governs how a model gets access to tools and data sources, not how an editor talks to an agent.

ACP in editors, products, and news coverage

You don’t see the protocol directly. You notice it by the fact that an editor suddenly offers several assistants to choose from. In Zed, for example, you can switch between different agents, including well-known command-line tools from major AI providers. Extensions that speak ACP also exist for other editors such as Neovim or Emacs.

In news reports, the term usually comes up in connection with standardization. Then the questions are things like: does a provider open up its assistant to third-party editors? A common misconception is that ACP is itself an AI. It is merely the agreement about the messages, comparable to a plug standard, which says nothing about the quality of the device.

For investors and observers, this topic is relevant because standards help decide market power. If an open protocol becomes established, the actual assistant becomes easier to swap out. Then what matters most is whose model delivers the best results, and less which program it is built into.