Three protocols, three different jobs. MCP plugs your agent into tools and data. A2A lets agents talk to other agents. AG-UI streams the agent's work to the humans watching it. The most common mistake is treating them as rivals — they're designed to layer. Answer 6 questions below and get your recommendation, with the security checklist for each protocol you pick.
Connects: your agent to tools and data — APIs, databases, files, services.
Think of it as: one universal port for every tool. The common analogy is "USB-C for agents."
Origin: Anthropic, November 2024; now governed by the Linux Foundation's Agentic AI Foundation (launched December 2025, co-founded by OpenAI, Anthropic, Google, Microsoft, AWS, and Block).
Scale (industry-reported): roughly 97 million monthly SDK downloads and 10,000+ production servers.
Use it when: your agent needs your CRM, your docs, or your code — anything it doesn't own.
Connects: agents to other agents — delegation and coordination across teams, vendors, and organizations.
Think of it as: a shared language between agents. The common analogy is "colleagues for agents."
Origin: Google, April 2025; v1.0 released March 2026 with signed Agent Cards; joined the Agentic AI Foundation in August 2026. IBM's ACP merged into A2A in August 2025.
Scale (industry-reported): 150+ organizations behind it.
Use it when: your agent hands work to another company's agent, or a fleet of agents must coordinate.
Connects: agents to the people using them, via event streaming.
Think of it as: a live feed between the agent and the app — people watch the agent work and approve steps as it goes.
Origin: CopilotKit and the open-source community.
Use it when: humans need to see what the agent is doing and sign off on steps — the human-in-the-loop layer.
A production agent system usually looks like this:
User —AG-UI→ Orchestrator —A2A→ Specialist agents —MCP→ Tools & data
The orchestrator delegates to specialists over A2A; each specialist calls tools over MCP; and everything streams to the human through AG-UI. As the saying goes: MCP gives your agent hands, A2A gives your agents colleagues — and AG-UI lets the humans supervise.
Honesty note: the comparison graphic that prompted this page claims MCP is near 500 million monthly SDK downloads; across industry sources the verifiable figure is roughly 97 million. This page uses the lower, verifiable number. All scale figures above are industry-reported, not independent findings.
MCP, A2A, and AG-UI solve different problems, and most teams only need one to start. Answer 6 questions about what your AI agents actually do, and this picker recommends the right protocol plus the security checklist for it.
No — they're complementary. MCP connects an agent to tools and data (vertical); A2A connects agents to other agents (horizontal). Production systems use both: the orchestrator delegates over A2A, specialists call tools over MCP.
MCP, in most cases. Connecting your agent to your own tools and data is the step nearly every project needs. A2A matters once you have multiple agents collaborating across teams or vendors.
AG-UI is the event-streaming protocol between agents and the interfaces people see — the live feed where humans watch the agent work and approve steps. It's the human-in-the-loop layer, complementary to both MCP and A2A.
Malicious servers hiding instructions in tool descriptions. Defenses: review every MCP server like a software dependency, grant least-privilege tools only, never expose a generic pass-through tool, and pin spec versions. Our open-source mcp-provenance-scanner checks tool descriptions for hidden instructions.
Agent Cards advertise capabilities publicly, so decide who may call what before production, require strong authentication for cross-organization calls, and log every delegation like a network call. Our open-source a2a-security-scanner audits Agent Card exposure.