Agent Client Protocol (ACP) Documentation
by Agent Client Protocol
The LSP for coding agents — one protocol so any editor can drive any agent, with official SDKs in five languages.
Overview
ACP standardizes JSON-RPC 2.0 communication between code editors and coding agents. It was created by Zed Industries and released in August 2025; the editor launches the agent as a subprocess over stdio, with streamable HTTP and WebSocket transports still in progress. The documentation is organized as Getting Started (introduction, architecture overview, agents directory, clients directory, and the ACP Registry launched January 2026); Protocol v1, which specifies initialization, authentication, session setup, session list, delete and config options, prompt turn, content blocks, tool calls, file system, terminals, agent plan, slash commands, session modes, elicitation, cancellation, extensibility, transports and the JSON schema; a Protocol v2 draft with a migration guide and a reworked prompt lifecycle; Libraries for Rust, TypeScript, Python, Java and Kotlin, with the Rust and TypeScript SDKs at v1.0; and an open RFD process with governance documents, working groups and OpenAPI specifications. It deliberately reuses MCP's JSON representations where possible and adds coding-specific types such as diffs — MCP connects an agent to tools, ACP connects an agent to an editor. Adoption is the reason to care: 37 agents implement it, including Gemini CLI, Codex CLI, Cursor, JetBrains Junie, Goose, OpenHands and GitHub Copilot in public preview, alongside Zed, JetBrains, VS Code, Neovim, Emacs and Obsidian on the client side. Apache-2.0, 3.9k GitHub stars.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Documentation
- Cost
- Free
- Duration
- ~3-4 hours to read the v1 protocol spec; a day or two to implement a minimal agent or client
- Provider
- Agent Client Protocol
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Model an editor-to-agent conversation as sessions and prompt turns over JSON-RPC
- ✓Implement initialization and capability negotiation between a conforming client and agent
- ✓Stream tool calls to an editor with permission prompts the user actually controls
- ✓Expose file system and terminal capabilities to an agent without handing over raw shell access
- ✓Render agent plans, diffs and slash commands using ACP's content block types
- ✓Handle cancellation, session resume, session list and session delete for long-running agents
- ✓Decide where MCP ends and ACP begins when wiring tools versus editor integration
- ✓Pick between the Rust, TypeScript, Python, Java and Kotlin SDKs for your integration
Highlights
- •Shipped adoption rather than a press release: 37 agents implement it, including Gemini CLI, Codex CLI, Cursor, JetBrains Junie and GitHub Copilot in public preview
- •Official SDKs in five languages, with Rust and TypeScript at v1.0, so you are not writing JSON-RPC plumbing by hand
- •Complements MCP rather than competing with it — it reuses MCP's JSON representations and adds editor-specific types like diffs
- •Governed in the open with an RFD process, working groups and published OpenAPI specs; a v2 draft and migration guide are already public
- •The ACP Registry, launched January 2026, turns wiring up a new agent into a config change rather than an integration project
Who It’s For
Best For
- ✓Editor and IDE authors who want to support many coding agents without one integration each
- ✓Teams building a coding agent that needs to reach users inside editors they do not own
- ✓Engineers deciding how MCP, A2A and ACP fit together in an agent architecture
Prerequisites
- •Comfort with JSON-RPC 2.0, subprocess I/O and asynchronous streaming
- •Familiarity with how coding agents work — tool calls, permission prompts and file edits
- •Helpful but not required: prior exposure to MCP or the Language Server Protocol
FAQ
What is Agent Client Protocol (ACP) Documentation?
The official specification and implementation guide for the Agent Client Protocol, a JSON-RPC standard that connects code editors and IDEs to coding agents. Aimed at engineers building either side of that boundary — an editor extension that needs to support many agents, or an agent that needs to reach users inside existing editors. After reading it you can implement a conforming client or agent against the published schema and SDKs.
Is Agent Client Protocol (ACP) Documentation free?
Agent Client Protocol (ACP) Documentation is free to access.
What level is Agent Client Protocol (ACP) Documentation for?
Agent Client Protocol (ACP) Documentation is aimed at a advanced audience. Recommended background: Comfort with JSON-RPC 2.0, subprocess I/O and asynchronous streaming, Familiarity with how coding agents work — tool calls, permission prompts and file edits, Helpful but not required: prior exposure to MCP or the Language Server Protocol.
How long does Agent Client Protocol (ACP) Documentation take?
Expect roughly ~3-4 hours to read the v1 protocol spec; a day or two to implement a minimal agent or client. Most learners work through it at their own pace.
What will I learn from Agent Client Protocol (ACP) Documentation?
You'll learn: Model an editor-to-agent conversation as sessions and prompt turns over JSON-RPC; Implement initialization and capability negotiation between a conforming client and agent; Stream tool calls to an editor with permission prompts the user actually controls; Expose file system and terminal capabilities to an agent without handing over raw shell access; Render agent plans, diffs and slash commands using ACP's content block types; Handle cancellation, session resume, session list and session delete for long-running agents; Decide where MCP ends and ACP begins when wiring tools versus editor integration; Pick between the Rust, TypeScript, Python, Java and Kotlin SDKs for your integration.
Topics
Sources
This page was written from 4 sources, 2 on domains other than agentclientprotocol.com.