This post, written in Chinese by Phodal Huang, starts from a problem he sees in enterprise development: several AI coding assistants — the post names Cursor, Copilot, Augment and Claude Code — now tend to appear in the same company, and platform teams must let them coexist across different editors without giving up efficiency, collaboration or governance. His answer is the Agent Client Protocol (ACP) (ACP), which his team had recently integrated into AutoDev, the AI coding tool they build.
He says his interest in ACP began with wanting to offer customers a unified AI coding agent service on an internal network. He contrasts enterprise work with vibe-coding demos: in an enterprise, people remain responsible for their own code and AI is only an assisting tool, so AI-written code still needs a tool for reviewing it. After finding that Claude Code’s IDEA plugin did not work well for this, the team turned to ACP as promoted by JetBrains.
Key Points
- ACP is presented as a standard protocol between IDEs or editors and AI coding assistants, which the author suggests thinking of as “LSP for AI agents”. Its core principle, as he puts it, is letting the IDE fully control every step the agent takes, by making tool calls first-class and visible to the IDE.
- The author describes its working model as the IDE acting as client and the agent as a server subprocess, communicating over JSON-RPC 2.0 on stdio, with capability negotiation at start-up, real-time updates on plans and tool calls during a session, and permission requests for sensitive operations such as editing files or running commands.
- AutoDev’s integration, made while building AutoDev 3.0 (Xiuper) around a unified rendering architecture, was bidirectional — AutoDev can act as an ACP server and as an ACP client — using the official Kotlin and TypeScript SDKs, and was built around a common renderer interface so that ACP events are rendered on multiple platforms; agents that do not speak ACP, such as Claude Code, are adapted by parsing their own streaming output.
- JetBrains, which the author calls one of ACP’s main promoters, offers an official ACP Agent Registry
for one-click installation and runtime management, custom configuration through an
acp.jsonfile, and MCP integration for agents running under ACP. - He summarises the relationship between the two protocols as ACP handling interaction between agent and IDE while Model Context Protocol extends the agent’s capabilities to external systems.
- He describes multi-agent workspaces that ACP makes easier to build: Augment Intent, which he had been invited to use, with a coordinator agent splitting tasks among specialist agents working in separate worktrees, and Google Antigravity with its editor view and agent manager surface.
- He proposes a three-layer governance model for enterprise AI platforms: MCP and Skills as the carrier of reusable enterprise capabilities, ACP as the interaction layer between developers and assistants, and Agent2Agent Protocol (A2A) as the collaboration network among agents.
Context
The post is a practitioner’s account of his own team’s integration work combined with a survey of vendors’ offerings; it reports no measurements. Its enterprise recommendations — that now is the best time to build a multi-agent development platform — rest on the author’s own view of where the ecosystem is heading.