Skip to content
Back to blog

MCP vs A2A vs ACP: Agent Tool and Agent Protocols

July 16, 20264 min read

Three patterns keep showing up in agent architecture slides: MCP (agent → tools), A2A (agent → agent via a registry of agent cards), and ACP (agent → agent over REST with a metadata manifest). They solve different connection problems — do not treat them as rivals for the same job.

This guide maps each flow, when to use it, and how they compose. MCP deep-dives live in [Nine AI Concepts for 2026](/blog/nine-ai-concepts-2026); multi-agent anatomy in [What Is Agentic AI?](/blog/what-is-agentic-ai).

Running example: a support agent that needs Postgres + GitHub (MCP), then delegates a specialist research agent (A2A or ACP) before answering.

MCPAgent → toolsHost · client · serverDB · API · filesA2AAgent → agentRegistry · cardsDelegate + clarifyACPAgent → agentManifest · RESTSync or async
Three protocols: tools (MCP) vs peers (A2A / ACP)

MCP: agent → tools

Model Context Protocol (MCP) connects an agent host to external tools and data. Flow: User request → MCP Host (chat UI, IDE like Cursor, or custom app) → MCP Client packages the need → MCP Server bridges to resources → tools (Postgres, GitHub APIs, Google Drive files) → structured response back up the chain.

MCP is not agent-to-agent messaging. It standardizes how one runtime discovers and calls capabilities — databases, APIs, files — without a custom plugin per model.

When to use: give one agent hands on your systems. When not to: peer delegation between two autonomous agents — that is A2A/ACP territory.

HostClientServerToolsResponse
MCP: User → Host → Client → Server → tools → structured response

Quick reference

  • Host examples: Claude desktop, Cursor/VS Code, custom AI apps.
  • Client routes; server exposes tools/resources with auth scopes.
  • Typical tools: DB, GitHub, Drive — structured results, not free-form HTML.
  • Keep REST for human clients; add MCP for agent-driven discovery.
  • Auth still matters — MCP is not a free pass to production data.
MCPPostgreSQLGitHubGoogle Drive

Remember this

I can trace MCP Host → Client → Server → tools → structured response.

A2A: agent → agent via registry

A2A (agent-to-agent) is peer discovery and delegation. Agent A gets the user request, queries a registry of Agent Cards (capability descriptions), finds Agent B, delegates the task, and consumes results. A dashed path covers “more information needed” so B can ask A for clarification.

The registry is the phone book; cards advertise what each peer can do. This fits multi-agent systems where specialists appear and disappear.

When to use: dynamic specialist discovery across a fleet. When not to: a fixed two-agent handoff you already hard-wire — a registry may be overhead.

11. Agent AUser request22. RegistryQuery agent cards33. CardFind capable peer14. Agent BExecute task25. ResultsBack to A36. Clarify?More info if needed
A2A: Agent A → registry (cards) → Agent B → results

Quick reference

  • Agent Card = capability metadata for a peer.
  • Registry returns cards; A delegates to B.
  • Results (and clarify loops) flow back to A.
  • Complementary to MCP: B may still use MCP for its own tools.
  • Industry experiments — design for versioned cards and auth between peers.

Remember this

I know A2A as registry + agent cards for peer discovery and task delegation.

ACP: agent → agent over REST

ACP frames agent-to-agent as REST/HTTP with a metadata manifest of available agents. Agent A reads the manifest, discovers Agent B, sends an HTTP request, and gets a sync or async response depending on the task.

If your org already operates APIs, gateways, and OpenAPI, ACP-style contracts feel familiar: agents are services with manifests instead of (or alongside) Agent Cards in a custom registry.

When to use: standardized HTTP between agent services in your mesh. When not to: IDE-local tool plugins — prefer MCP for host↔tool.

11. Agent ANeed a peer22. ManifestRead capabilities33. Discover BPick agent14. REST callHTTP request25. Sync / AsyncBy task length36. CompleteA continues
ACP: Manifest discovery → REST/HTTP → sync or async result

Quick reference

  • Manifest lists capabilities (like a service catalog).
  • Communication: REST/HTTP request/response.
  • Sync for quick tasks; async when work is long-running.
  • Fits API gateways, mTLS, and existing observability.
  • Still need schemas for task payloads and errors.

Remember this

I can describe ACP as manifest discovery plus REST sync/async between agents.

How to choose — and combine them

Use the connection type as the filter: agent → toolsMCP. agent → agent with a capability registry/cards → A2A. agent → agent with HTTP + manifest and sync/async semantics → ACP.

Real systems compose: the host uses MCP for Postgres/GitHub; the orchestrator A2A-delegates a research peer; that peer might expose ACP endpoints to other services. Start with MCP for tool access; add A2A/ACP only when you truly need peer agents.

What are you connecting?Tools / data→ MCPPeers + cards→ A2APeers + HTTP→ ACPComposeMCP + A2A/ACP
Zoom: pick by connection type — tools vs peers

Quick reference

  • MCP = tools/data plane for one agent runtime.
  • A2A = peer discovery via registry + cards.
  • ACP = peer calls via REST + manifest.
  • Do not replace MCP with A2A for database access.
  • Practice: draw your support bot — label each hop MCP, A2A, or ACP.

Remember this

I can pick MCP vs A2A vs ACP from tools vs peer-agent needs — and combine them.

Key takeaway

Share:

MCP gives agents tools. A2A and ACP give agents peers — one via registry cards, one via REST manifests. Confusing the three produces either a mega-prompt with no real integrations or a mesh of agents that cannot touch a database.

Your homework: for one product flow, list every hop. Tag each MCP / A2A / ACP / plain REST. Delete any hop that does not earn its protocol. Then implement one MCP server before standing up a registry.

Related Articles

Artificial intelligence is not here to replace .NET developers. It is here to extend what they can build. The fundamenta

Read

An LLM alone is clever autocomplete — it predicts the next token, frozen at training time, with no hands and no durable

Read

Clever prompts help. They do not replace production AI design. In 2026 the useful concepts are loops with feedback, tool

Read

Keep learning

Follow a structured path or browse all courses to go deeper.