Agent Client Protocol (ACP): When and How to Use It
ACP is the open standard that lets any coding agent work in any supporting editor — built by Zed and JetBrains in Oct 2025. This article covers when ACP adds value over built-in IDE AI, how the protocol actually works, how to build an ACP-compatible agent with deepagents-acp, and what breaks in production.
Quick Reference
- →ACP = Agent Client Protocol — open standard, created by Zed + JetBrains (Oct 2025)
- →Transport: JSON-RPC 2.0 over stdio (local); HTTP/WebSocket for remote (WIP)
- →Editors: Zed and JetBrains natively; VS Code and Neovim via community extensions
- →Distinct from MCP (tools) and A2A (agent-to-agent) — ACP is specifically for IDE integration
- →Capabilities negotiation: IDE and agent declare what each can offer before any task runs
- →Agent Registry (Jan 2026): searchable catalog of ACP agents built into JetBrains + Zed
- →Security: ACP agents inherit the editor's filesystem access — scope carefully
Should You Use ACP?
The answer depends on whether your team's coding agent needs identity across editors, or just access to one editor's built-in AI. Built-in IDE AI (GitHub Copilot in VS Code, JetBrains AI Assistant, Cursor) is the right default for most teams — zero configuration, deeply integrated, maintained by the editor vendor. ACP becomes worth the investment when two or more of these are true:
- ▸Your team uses multiple editors (some on Zed, some on JetBrains) and you want one agent with consistent behavior across all of them
- ▸You need a domain-specific agent trained on your codebase conventions, internal APIs, or proprietary tooling that built-in AI doesn't know
- ▸You want your agent to call MCP tool servers (databases, CI systems, internal APIs) — ACP lets the agent negotiate which MCP endpoints the IDE should pass
- ▸You're building an agent that needs to be auditable and company-controlled, not dependent on a SaaS AI product's availability or data policies
- ▸You want to use a specific model or provider that isn't available in built-in IDE AI
Don't build an ACP agent if your team uses a single editor and built-in AI already handles your use cases. ACP adds a Python subprocess, a capabilities handshake, and a custom agent to maintain. The investment pays off at scale or with custom tooling — not for general coding assistance that Copilot already does well.
| Situation | Use built-in IDE AI | Use custom ACP agent |
|---|---|---|
| General coding, refactoring, docs | ✓ Best default | Overkill |
| Team uses Zed + JetBrains + VS Code | Inconsistent per-editor | ✓ One agent, all editors |
| Need internal APIs / proprietary tools | ✗ Not possible | ✓ Via MCP + ACP |
| Company controls model and data | Depends on vendor | ✓ Full control |
| Team-specific conventions and memory | Generic suggestions | ✓ Domain-specific |
| You want to ship in 10 minutes | ✓ Zero config | Build time required |