ACP
ACP (Agent Client Protocol): an open protocol created by Zed in August 2025 that standardizes how code editors talk to AI agents, regardless of model or vendor.

ACP (Agent Client Protocol) is an open protocol that standardizes communication between a code editor and an AI agent. Created by Zed Industries in August 2025 and released under the Apache license, it complements MCP by operating at a different layer of the agentic stack.
Technical definition
The combinatorial problem ACP solves
Without ACP, every editor (VS Code, JetBrains, Zed, Cursor, Neovim) has to integrate every AI agent (Claude Code, Gemini CLI, Codex CLI, Mistral Vibe, etc.) by hand. That's N × M integrations to maintain — exactly the problem the Language Server Protocol (LSP) solved for programming languages in the 2010s.
ACP applies the same logic to AI agents:
- The ACP client (the code editor) speaks ACP
- The ACP server (the AI agent) exposes its permission requests, file writes and shell commands through the protocol
- The editor can intercept, display and arbitrate every agent action in a unified UI
ACP vs MCP
The two protocols are complementary, not competing:
| Aspect | MCP | ACP |
|---|---|---|
| Creator | Anthropic | Zed Industries |
| Date | November 2024 | August 2025 |
| Direction | Agent → tools/data | Editor → agent |
| Typical use case | Connecting an agent to Figma, GitHub or a SQL database | Showing a Claude Code agent's permission requests inside VS Code |
| Transport | JSON-RPC 2.0 (stdio, HTTP, WebSocket) | JSON-RPC 2.0 (stdin/stdout) |
A mature AI agent setup in 2026 uses MCP to extend capabilities and ACP to govern the user interface.
Current challenges
Adoption in 2026
When Zed launched it in August 2025, ACP was primarily an internal integration mechanism. JetBrains joined shortly after, turning ACP into a broader interoperability layer for agentic coding workflows. Gemini CLI is supported natively; Codex CLI goes through an adapter maintained by the ACP project, and Claude Code through a bridge adapter built by Zed (official list of ACP agents). As of November 2026, Anthropic still has not adopted ACP natively.
What it changes for a CDO
For a decision-maker, ACP means AI agents stop being locked into their original editor. A team can adopt Claude Code today, switch to Gemini CLI tomorrow and plug in an in-house agent the day after — without rewriting its editor tooling. It's the same logic MCP brought to the tools side: lower vendor lock-in risk, faster rollout.
ACP also plays an underrated governance role: by intercepting 100% of the JSON-RPC RequestPermissionRequest messages flowing between the agent and the user, ACP is the ideal layer for plugging in a physical approval mechanism or a hierarchical escalation workflow. That's the step forward compared with MCP, where the agent remains free to call an approval tool or not.
Known limitations
- No native Anthropic support so far — Claude Code goes through a bridge adapter, which adds a layer of indirection.
- Gradual mainstream adoption: VS Code has not (yet) announced a native implementation — which fueled the competitive reading of "Zed and Google versus VS Code lock-in" in the trade press.
- Young protocol: stable version 1 dates from late 2025, and the ecosystem is still taking shape.
Frequently Asked Questions
What is the Agent Client Protocol (ACP)?
ACP is an open protocol created by Zed Industries in August 2025 that standardizes communication between a code editor and an AI agent. It uses JSON-RPC 2.0 over stdin/stdout — like MCP, but in the opposite direction: MCP lets agents call tools; ACP lets editors drive and observe agents. Open source under the Apache license, ACP has reached stable version 1.
What is the difference between ACP and MCP?
The two protocols are complementary, not competing. MCP (Model Context Protocol, Anthropic) connects the AI agent to external tools and data — the server exposes capabilities, the agent uses them. ACP (Agent Client Protocol, Zed) connects the code editor to the AI agent — the editor intercepts the agent's requests (permissions, file writes, shell executions) and presents the UI to the user. A robust setup uses both side by side.
Which editors and agents support ACP?
On the editor side: Zed natively, and JetBrains joined the protocol shortly after launch. On the agent side: Gemini CLI is supported natively, Codex CLI goes through an adapter maintained by the ACP project and Claude Code through a bridge adapter built by Zed (official list of agents). As of November 2026, Anthropic has not adopted ACP natively. An official ACP registry lists validated implementations.
Why was ACP created?
To solve the N × M combinatorial problem: N editors × M AI agents = N × M bespoke integrations to maintain. ACP turns this into N + M: each editor implements ACP once, each agent implements ACP once, and everything plugs together. It is exactly the logic that made Microsoft's LSP (Language Server Protocol) a success for programming languages in editors.