Agent Client Protocol: Because Apparently We’re Still Writing Custom Integrations in 2025 🙃

So… you like writing one-off plugins for every AI tool + editor combo? No? Cool, neither do I.

Guess what? Agent Client Protocol (ACP) exists so we can stop reinventing the wheel every time a new AI agent drops. Shocking, I know.


The Problem Nobody Asked For (But We All Have)

Every editor + AI combo today:

  • VS Code + Copilot → works
  • Neovim + “random AI agent” → lol good luck
  • PhpStorm + anything → 👴 grandfathered chaos

Everyone’s out here writing custom adapters, like it’s 2012 and SOAP APIs are still cool. ACP’s answer: “what if we, like, didn’t?”.


“AI integrations are hard” — developers, 2025
“Just follow the protocol” — ACP, probably

So, What Is ACP Anyway?

TL;DR: ACP is like the LSP (Language Server Protocol) but for AI agents.

You want your editor to talk to your AI agent? Use ACP. Done.

How it works:

  • Agent = the AI thingy (e.g., Gemini CLI, Claude, whatever OpenAI’s cooking this week)
  • Client = your code editor (Zed, Neovim, maybe VS Code if Microsoft ever stops gatekeeping)
  • JSON-RPC 2.0 = the lingua franca between them
  • Markdown = default response format because, apparently, everyone loves rendering tables now

Mechanics:

  • Editor launches the agent as a subprocess
  • They chat via stdin/stdout with JSON messages
  • Agents can stream partial responses so you get ✨ live ✨ “thoughts”
  • Agents can request file writes, tool execution, etc.
  • Editor mediates these calls and can ask you: “bro, you sure you want this agent rm -rf /?”

ACP lets you bring your own AI agent without duct-taping plugins together. Revolutionary, I know.


Why You Should Care (Especially if You Touch Magento 2)

AI agents are increasingly being used to:

  • Refactor PHP codebases nobody wants to touch
  • Generate integration tests at 3 AM
  • Summarize controller logic so you can pretend you read it

Without ACP, you’re stuck hoping your AI tool even supports your editor. With ACP, you just… switch agents without switching tools. It’s like swapping npm packages — except hopefully less cursed.


Security Vibes™

Okay, so, AI agents touching your filesystem. What could possibly go wrong?

ACP assumes your agent is trusted (lol, okay). But:

  • Editor mediates every destructive action
  • User gets final say on “pls delete prod DB” moments
  • Everything stays local unless the agent explicitly calls out to external APIs

Translation: ACP is “secure” as long as you don’t install sketchy AI agents from npm install totally-legit-ai@latest.


The State of ACP Right Now

  • Zed: First editor with ACP baked in
  • Gemini CLI: Reference AI agent
  • Neovim: Community plugin exists (CodeCompanion)
  • VS Code: lol no, Microsoft wants you in Copilot jail
  • PhpStorm: …look, we’re manifesting

ACP has TypeScript + Rust libs, so you don’t have to YOLO your own implementation.


Action Items for Your Sanity

  • Try it → Install Zed + Gemini CLI → vibe check ACP IRL
  • Watch adoption → Keep an eye on plugins for your editor
  • Write once, integrate everywhere → If you’re building an AI tool, ACP saves you from writing 17 editor-specific adapters
  • Don’t trust random agents → Self-explanatory, hopefully

Key Takeaways (a.k.a. TL;DR for ADHD Devs)

  • Stop writing custom AI-editor integrations → use ACP
  • JSON-RPC everywhere, Markdown always, diff display included
  • Local-first, editor-mediated, user-controlled
  • Already works in Zed + Neovim, Gemini CLI ships as reference
  • VS Code support? Ha. Don’t hold your breath.

References