ARD vs. Agent Registry vs. MCP vs. A2A
On this page
Quick answer
ARD is a federated description and search specification. An agent registry is a catalog service and policy boundary. MCP describes how a client connects to servers and tools. A2A describes agent-to-agent interaction. These layers can complement one another; none automatically replaces the others.
ARD can return an entry whose artifact is an MCP Server Card or A2A Agent Card. The client then follows that artifact’s protocol and local policy to authenticate and interact with the resource.
Responsibility map
| Layer | Primary job | Does not automatically provide |
|---|---|---|
| ARD | Describe, publish, index, filter, search, and federate agentic resource entries | Resource authentication, installation, invocation, approval, or action authorization |
| Agent registry | Operate a catalog with local ingestion, curation, access, approval, lifecycle, and search policy | Cross-environment interoperability unless it implements a shared interface |
| MCP | Connect an AI client to MCP servers, resources, prompts, and tools | Global discovery, publisher approval, or permission to execute every tool |
| A2A | Describe and coordinate tasks between agents | Global catalog federation or enterprise admission policy |
The current ARD v0.91 Proposal is artifact-agnostic: its type points to the underlying format while the artifact protocol owns its internal schema.
ARD and agent registries
ARD defines a common REST discovery baseline so different registries can participate in federation. A registry remains an implementation that chooses what to ingest, how to rank, who can search, what to approve, and when to revoke.
AWS describes ARD as an open standard—not a product or single registry—and as a potential complement to AWS Agent Registry. The AWS service announcement does not prove that every Agent Registry deployment currently implements v0.91 ARD federation. Verify live product documentation and endpoint behavior.
Use the ARD federation and ranking guide to define curation and trust.
ARD and MCP
ARD may describe an MCP server through an artifact entry and can expose search through an optional MCP wrapper. These are two separate uses:
- MCP server as a discovered resource: ARD returns a description that points to the server card.
- MCP as a discovery client wrapper: A chatbot calls an Agent Finder through an MCP tool.
In either case, discovery does not grant tool access. The MCP client still must trust the server, authenticate, review tool schemas, apply user and data permissions, confirm sensitive actions, cap cost, log calls, and support revocation.
The v0.91 proposal notes that full media-type registration for some MCP and A2A card types is still pending and formats may change. Avoid rigid intermediaries that assume final registration.
ARD and A2A
ARD can describe an A2A agent card so it can be found by task, capability, publisher, or filters. A2A then governs the actual agent communication pattern.
A relevant agent is not necessarily an approved peer. Verify identity, endpoint, supported protocol/version, skills, authentication, data policy, task authority, callbacks, artifacts, timeouts, cost, monitoring, and cancellation before sending work.
The complete lifecycle
publish description
-> ingest and validate
-> search and rank
-> verify publisher and artifact
-> curate and approve
-> install or connect
-> authenticate
-> authorize data and action
-> invoke through MCP, A2A, API, skill, or workflow
-> monitor, review, revoke, and recoverARD primarily covers the first three steps and supplies metadata for later decisions. A registry can cover ingestion through approval. The artifact protocol covers connection and interaction. Your governance system must join the evidence.
Avoid these category errors
- A search score is not a security rating.
- A verified publisher is not proof of safe code.
- A catalog approval is not runtime authorization for every user.
- An MCP connection is not approval for every tool call.
- An A2A agent card is not proof that the peer will complete a task correctly.
- A public manifest is not consent to install or send data.
- A common description does not guarantee client compatibility.
Choose the layer by problem
Use ARD when the problem is cross-registry resource discovery. Use an agent registry when the problem is operating a governed catalog. Use MCP when an AI client needs a standardized server/tool connection. Use A2A when one agent needs to delegate or coordinate with another agent.
Many production systems need all four, plus identity, secrets, policy, approval, audit, cost, monitoring, incident response, and recovery.
Run the ARD agent catalog readiness check before joining discovery to runtime execution.
Frequently asked questions
Is ARD a competitor to MCP?
Not primarily. ARD focuses on finding resources; MCP focuses on connecting clients to MCP servers and tools. ARD can discover MCP artifacts or be exposed through an MCP wrapper.
Is AWS Agent Registry the same as ARD?
No. AWS describes Agent Registry as a product catalog and ARD as an open federation standard that may complement it. Verify current implementation support separately.
Does ARD authenticate agents?
No. The specification delegates resource authentication to the underlying artifact protocol and runtime.
Can a client invoke the first search result automatically?
Technically an implementation could, but relevance alone is insufficient. Add publisher verification, local approval, artifact validation, authentication, least privilege, action policy, monitoring, and recovery.
Official sources
Source check: August 27, 2026. Recheck ARD version, media types, protocol wrappers, registry products, MCP/A2A specifications, and implementation support.