Skip to content

AI Agent Protocols

This document describes the key protocols that enable interoperability between AI agents, tools, and development environments.

1. Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open standard that standardizes how applications interact with large language models (LLMs) and provide them with tools and resources.

  • Developer: Anthropic
  • Purpose: Decouples the "brain" (LLM) from the "tools" (APIs, databases, local files).
  • Key Concepts:
    • MCP Servers: Host specific tools (e.g., Google Calendar, GitHub, ClickHouse).
    • MCP Clients: Frameworks or IDEs that connect to servers to use their tools (e.g., Claude Agent SDK, Zed, Cursor).
  • Benefits: Build a tool once as an MCP server and use it in any compatible agent framework or editor.
  • Sources: Making MCP cheaper via CLI (Exploring lightweight CLI implementations vs server-side MCP).

2. Agent Client Protocol (ACP)

The Agent Client Protocol (ACP) is an open standard designed to enable any AI agent to integrate seamlessly with any code editor or editing environment.

  • Developer: Zed
  • Purpose: Standardizes the interface between terminal-based or external agents and the IDE's UI components (multi-file editing, syntax highlighting, diff viewing).
  • Key Concepts:
    • Universal Compatibility: Any agent implementing ACP can gain access to an IDE's full codebase context and powerful reviewing tools.
    • Privacy First: ACP is designed to be local-first; data doesn't necessarily touch cloud servers unless specifically configured.
  • Benefits: Allows developers to use specialized external agents (like Claude Code or Gemini CLI) directly inside their preferred IDE without proprietary plugins for each agent.