ACP: A Universal Protocol for AI Agent Clients

The Case for Agent Interoperability

Most current AI agent harnesses rely on bespoke, proprietary interfaces that tie a specific agent to a single client application. This fragmentation mirrors a hypothetical web where every website required a unique browser. Alex Hancock argues that while the Model Context Protocol (MCP) successfully standardized how agents interact with external tools, the industry lacks a corresponding standard for how client software (such as text editors or terminal interfaces) communicates with agent harnesses to assign tasks and receive status updates.

The Agent Client Protocol (ACP)

ACP is a JSON-RPC based standard designed to decouple client applications from agent harnesses. By providing a common language for sessions, user messages, tool call notifications, and permission requests, it allows a single client implementation to drive any compliant harness. The protocol is intentionally extensible, allowing developers to implement custom methods prefixed with an underscore. This design pattern enables the community to identify common operational requirements through real-world usage before promoting them to the official standards track.

Enabling Remote Agent Architectures

While initial implementations focused on local standard-IO communication, the protocol now includes specifications for HTTP and WebSocket transports. This shift allows the four primary components of the agentic stack—the client, the harness, the tools (via MCP), and the model—to be independently deployed across different machines or cloud environments. This modularity enables developers to build specialized, white-labeled, or domain-specific clients that remain compatible with any ACP-compliant harness, fostering a competitive market for client-side user experience.