Developer communities have been hearing two protocol names a lot lately: MCP (Model Context Protocol) and ACP (Agent Communication Protocol). Both are open standards, but they serve different purposes. Teams that pick the wrong layer can't scale.
MCP — Model ↔ Product
MCP standardizes how an AI model (or agent) holds a structured conversation with a product. Anthropic released it in 2024; it parallels OpenAI Actions but is more flexible.
- Transport: HTTP or stdio.
- Three core concepts: tools, resources, prompts.
- Discovery:
.well-known/mcp.json. - Used by: Claude Desktop, Cursor IDE, ChatGPT MCP plug-ins.
Detailed guide: What is MCP?.
ACP — Agent ↔ Agent
ACP standardizes how multiple agents exchange structured messages with each other. A draft initiative led by IBM, Cisco, and others.
- Transport: HTTP or WebSocket.
- Core concepts: agent profile, message envelope, capability discovery.
- Discovery:
.well-known/acp.json(still in draft). - Use cases: multi-agent workflows (orchestrator + specialist agents).
When should you use which?
| Scenario | Right protocol | | --- | --- | | My product will be triggered from Claude Desktop | MCP | | My agent needs to negotiate with another agent | ACP | | Must be reachable from both ChatGPT and Claude | MCP | | Microservice architecture with "agent-strong" workflows | ACP |
Practical recommendation
For most SMBs and SaaS, MCP is the right first step. ACP comes in as an additional layer once you progress to advanced orchestration.
Further reading
For your integration strategy, see our API documentation and contact page.
