Technology RadarTechnology Radar

Agent Client Protocol (ACP)

protocolagentopen-source
This item was not updated in last three versions of the Radar. Should it have appeared in one of the more recent editions, there is a good chance it remains pertinent. However, if the item dates back further, its relevance may have diminished and our current evaluation could vary. Regrettably, our capacity to consistently revisit items from past Radar editions is limited.
Trial

The Agent Client Protocol (ACP) is an open standard that defines how code editors communicate with AI coding agent — the same problem that LSP solved for language intelligence. Created by Zed Industries at Google's request, co-developed with JetBrains, and now supported by 30+ agent including Gemini CLI, Claude Code, GitHub Copilot CLI, Goose, and Kiro.

The Problem It Solves

Without ACP, every new agent-editor combination requires custom integration work. Agents must implement editor-specific APIs, and editors must build bespoke integrations for each agent. This is the N × M problem — the same one that LSP (Language Server Protocol) solved for language intelligence a decade ago.

ACP reduces this to N + M: agent implement one protocol, editors adopt one protocol, and everything works together.

How It Works

ACP uses JSON-RPC 2.0 over stdio (local agent) or HTTP/WebSocket (remote agent) to relay user requests to agent and render responses. The protocol defines a minimal set of endpoints:

Endpoint Purpose
session/new Create a new agent session
session/prompt Send a user message to the agent
session/load Resume a previous session
Permission requests Agent asks the editor for approval before tool execution
File operations Agent reads/writes files through the editor

Where It Fits in the Stack

ACP, MCP, and A2A operate at different layers and are complementary:

Editor <──ACP──> Agent <──MCP──> Tools/Data
                   │
                  A2A
                   │
                 Other Agents
  • ACP standardises editor-to-agent communication (this entry)
  • MCP standardises agent-to-tool communication
  • A2A standardises agent-to-agent communication

Editor Support

Editor Status
Zed Native (first-party)
JetBrains IDEs Integrated via AI Assistant plugin; co-developed with Zed
Neovim Via CodeCompanion and avante.nvim plugins
Emacs Via agent-shell plugin
Obsidian Via community plugin
VS Code Not supportedopen issue under discussion, but no commitment from Microsoft

The VS Code gap is significant — it holds ~70% editor market share. However, Microsoft is adopting ACP on the agent side (Copilot CLI) while pursuing its own approach (AgentHQ) on the editor side.

Agent Support

30+ agent are listed in the ACP Registry (launched January 2026):

Agent Provider Notes
Gemini CLI Google Reference implementation
GitHub Copilot CLI Microsoft/GitHub GA since February 2026
Claude Code Anthropic Via claude-agent-acp adapter (not native)
Goose Block Native ACP
Kiro CLI AWS Native ACP
Codex CLI OpenAI Via adapter
Mistral Vibe Mistral Native ACP
Kimi CLI Moonshot AI Native ACP
Cline Community Supported
Qwen Code Alibaba Supported

Why It's in Trial

For Trial:

  • JetBrains co-developing the protocol and integrating across their entire IDE lineup — the second-largest IDE vendor after Microsoft
  • Google initiated the project and contributed the reference implementation (Gemini CLI)
  • GitHub Copilot CLI reached GA with ACP support in February 2026
  • SDKs in 4 languages (Python, Rust, TypeScript, Kotlin)
  • 30+ agent in the registry, including offerings from Google, Microsoft, Anthropic, AWS, Block, and OpenAI

Why not Adopt:

  • Pre-1.0 spec (v0.11.3 as of March 2026) — still evolving, expect breaking changes
  • VS Code holdout — the dominant editor hasn't committed, limiting reach
  • Remote agent support incomplete — cloud/remote deployment scenarios still being worked on
  • Claude Code requires an adapter, not native support — Zed maintains the bridge
  • No standardised permission model — permission handling is inconsistent across implementations

Key Characteristics

Property Value
Created by Zed Industries (at Google's request)
Co-developed with JetBrains (since October 2025)
Announced August 27, 2025
Current version v0.11.3 (March 18, 2026)
License Apache 2.0
Transport JSON-RPC 2.0 over stdio / HTTP / WebSocket
Registry agentclientprotocol.com
GitHub agentclientprotocol/agent-client-protocol

Further Reading