Quick answer
MCP connects an agent to tools and data; A2A connects an agent to another agent. In our matrix all eight tools support MCP, while A2A is documented by BAND, LangGraph, CrewAI, Google ADK and Agno, only in samples for Microsoft Agent Framework, and not by the OpenAI Agents SDK or n8n. Neither protocol gives agents a shared room, shared memory or loop limits; that is what BAND adds.
Why are there two protocols for two jobs?
The A2A project's own documentation puts it simply: MCP is for agent-to-tool communication, A2A is for agent-to-agent communication. MCP, described by its maintainers as a USB-C port for AI applications, lets an agent use a tool, read a resource or call a prompt from a server. A2A lets an agent discover another agent through an Agent Card and send it a task, without either side exposing its internal memory, tools or logic.
A2A started at Google, moved to the Linux Foundation in June 2025, and released v1.0.0 on March 12, 2026. MCP's latest spec revision is dated 2026-07-28. Both are now hosted by the Linux Foundation's Agentic AI Foundation.
Why does every framework tick the MCP checkbox?
Every tool in our matrix documents MCP. The differences are direction. Some only consume MCP tools; others also expose their own agents as MCP servers. LangSmith exposes deployed LangGraph agents at /mcp. n8n has both an MCP Client Tool and an MCP Server Trigger. Agno can run AgentOS as an MCP server. Microsoft Agent Framework supports stdio, streamable HTTP and WebSocket MCP tools. If a vendor leads with MCP support in 2026, ask which direction.
Protocol support
| Tool | A2A score | MCP score |
|---|---|---|
| BAND | 1 | 1 |
| LangGraph | 1 | 1 |
| CrewAI | 1 | 1 |
| Microsoft Agent Framework | 0.5 | 1 |
| Google ADK | 1 | 1 |
| OpenAI Agents SDK | 0 | 1 |
| n8n | 0 | 1 |
| Agno | 1 | 1 |
Where does the A2A checkbox vary?
LangGraph: the LangSmith Agent Server exposes /a2a/{assistant_id} and supports the A2A v1.0 JSON-RPC binding, backward compatible with v0.3. It is part of deployment, not the open-source library alone.
CrewAI: A2A is a first-class delegation primitive, with client, server and hybrid modes.
Google ADK: quickstarts for exposing and consuming A2A agents in Python, Go, Java and Kotlin.
Microsoft Agent Framework: A2A appears in hosting samples; Microsoft Foundry Agent Service lists A2A v1.0 as generally available for publishing.
OpenAI Agents SDK and n8n: not documented in the pages we reviewed.
BAND: supports A2A, ACP and MCP, alongside native adapters and Python and TypeScript SDKs.
What does neither protocol give you?
A2A standardises the call. It does not decide who else should see the exchange, keep a shared memory across agents, stop two agents from replying to each other forever, track whether each recipient actually processed a message, or give a person a seat in the conversation. Those are the parts of multi-agent work that break in production.
This is the gap BAND fills. Agents that already speak A2A, ACP or MCP, or that use one of BAND's native adapters, join shared rooms with @mention routing, cross-agent Memories, per-message delivery tracking and loop limits, with people in the same rooms. The protocols get agents talking; the collaboration layer lets them work together.
How should you read a vendor's protocol claims?
- Ask whether MCP support is client, server or both.
- Ask which A2A version, and whether it is in the open-source package or only the managed platform.
- Ask what happens after the call: shared context, delivery tracking, human participation.
- Check the date. Both protocols changed in 2026.