MCP vs. A2A: What you Need to Know

Google announces support for MCP and announces A2A

MCP has exploded onto the scene in the last 6 months, and just when we see integration upon integration, on the first day of Google Cloud Next Sundar answered his question in the only way possible... ish!

The Gemini team announced they are supporting it, and asked the community for thoughts on what they want to see:

Then, Google rebranded their "Gen AI Toolbox for Databases" to MCP Toolbox for Databases.

But then we saw a different kinda of protocol from Google announced at Next, the Agent 2 Agent (A2A) protocol. This created some confusion, as it appears to have some overlap with MCP, and Google was careful to position it as complementary with a throw away line in their post:

"A2A is an open protocol that complements Anthropic's Model Context Protocol (MCP), which provides helpful tools and context to agents."

Dig a lil deeper though and you see them make the case for how the protocols work at different levels of abstraction:

Model Context Protocol (MCP) is the emerging standard for connecting LLMs with data, resources, and tools. We already observe MCP standardizing ‘function calling’ across different models and frameworks. This is creating an ecosystem of tool service providers and dramatically lowering the complexity to connect agents with tools and data. We expect this trend to continue as more frameworks, service providers, and platforms adopt MCP.

A2A is focused on a different problem. A2A is an application level protocol that enables agents to collaborate in their natural modalities. It allows agents to communicate as agents (or as users) instead of as tools. We hope that A2A gains adoption as a complement to MCP that enables ecosystems of agents and will be working in the open with the community to make this happen.

Getting back to the announcement, what sticks out to you when you read it? For one, as fits with Google Cloud Next, it is a very Enterprise-y announcement. Compared to the grass roots, bottoms up adoption of MCP, A2A comes out of the gate with company "partnerships" such as Accenture and SAP and Deloitte...

The whole vibe is: “Will this become the modern WSDL compared to MCP's simple JSON over REST?” #iykyk

Speaking of REST and MCP... what if you could take any of your OpenAPI REST endpoints and turn them into MCP resources automatically? The legendary Ray Cromwell just open sourced hippycampus, which does just that. The MCP creators at Anthropic have been talking about how they want to see more MCP resources being used vs. the more used tools, so hopefully this will take away a lot of the friction!

It's not all rose-y in MCP land though. Simon Willison, LLM prompt injection drummer, wrote up his latest warning around this as it relates to MCP. We have to be very careful as both users and developers, but we also really need more deep thinking here and better solutions.

Back to the original question at hand… we all know that people are still going to MCP, so the only way out is probably through!