Join our Newsletter — 33% off our NHI Course
Home› Guides› MCP vs A2A: How the Two Agent Protocols…
Guide AI Platforms

MCP vs A2A: How the Two Agent Protocols Differ and Fit Together

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, edited by Lalit Choda Updated 15 November 2026 12 min read
On this page

MCP (Model Context Protocol) is an open protocol that connects an AI application to tools, data and prompts, while A2A (Agent2Agent) is an open protocol that lets independent AI agents discover each other and work together on tasks. Put simply, MCP gives one agent more capabilities and A2A lets that agent hand work to other agents. Both are now projects of the Agentic AI Foundation under the Linux Foundation, and the A2A project itself describes the two as complementary rather than competing. This explainer sets out what each protocol defines, where they overlap, how they combine in one system, and what changes for identity and security when you use both. It is for architects choosing an integration pattern, developers building agents, and security teams reviewing agent designs.

Last checked: 2026-11-15. Protocols are listed alphabetically where compared (A2A, then MCP); order is never a ranking.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

  • Different relationships. MCP models an agent using a capability (agent to tool or data); A2A models agents collaborating as peers (agent to agent), including agents run by other teams or organisations.
  • Vertical and horizontal. The A2A documentation calls MCP the vertical layer that deepens a single agent and A2A the horizontal layer that connects agents across boundaries. Most production systems that use both apply A2A between agents and MCP inside each one.
  • Opaque versus exposed. A2A agents collaborate without sharing internal state, memory or tools. MCP servers deliberately expose their tools, resources and prompts to the client.
  • Both are stable and both moved recently. A2A reached v1.0 on 12 March 2026 (v1.0.1 on 26 May 2026). MCP's current specification revision is 2026-07-28, which made the protocol stateless and deprecated several features.
  • Two protocols, two trust boundaries. MCP authorisation is OAuth 2.1 based and optional; A2A declares security schemes in a signed or unsigned Agent Card. Each connection is a non-human identity to govern.

What each protocol is

A2A (Agent2Agent)

The A2A specification describes an open standard for communication and interoperability between independent, potentially opaque AI agent systems. Its stated goals are interoperability, collaboration, discovery, flexibility (request and response, streaming and push notifications), security and asynchronicity, including long-running tasks with a human in the loop. The project was contributed by Google, is licensed under Apache 2.0, is governed by a Technical Steering Committee with nine company seats, and was accepted as a Growth Stage project at the Agentic AI Foundation, announced on 27 August 2026.

MCP (Model Context Protocol)

The MCP specification defines three roles: hosts (the LLM application), clients (connectors inside the host) and servers (which provide context and capabilities). Servers offer resources, prompts and tools; clients can offer features such as elicitation back to servers. Messages are JSON-RPC 2.0. Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation, announced on 9 December 2025, and the project states that its maintainers keep control of technical direction. The project is moving its licence from MIT to Apache-2.0, with new code and specification contributions under Apache-2.0. For a full introduction, see What Is MCP? The Model Context Protocol Explained.

Side-by-side comparison

As of 15 November 2026, from each project's specification and repository.

FieldA2AMCP
PurposeAgents discover each other, negotiate interaction modes, manage shared tasks and exchange contextAI applications connect to external tools, data sources and prompt templates
Relationship modelledPeer agent to peer agent (client agent and remote agent)Host and client to server
Core unitsAgent Card, Task, Message, Part, ArtifactTools, resources, prompts; client features such as elicitation
Wire formatsThree bindings with functional equivalence: JSON-RPC, gRPC and HTTP+JSON/REST; streaming over Server-Sent EventsJSON-RPC 2.0 over stdio or Streamable HTTP
DiscoveryAgent Card at /.well-known/agent-card.json, registries or catalogues, or direct configurationserver/discover RPC (2026-07-28), client configuration, and the MCP Registry
StateStateful tasks with lifecycle states such as TASK_STATE_WORKING, TASK_STATE_INPUT_REQUIRED and TASK_STATE_AUTH_REQUIREDStateless protocol since 2026-07-28; servers that need state pass explicit handles as tool arguments
Long-running workNative: polling, streaming subscriptions and push notifications to webhooksTasks extension (io.modelcontextprotocol/tasks), opt-in
Authentication modelSecurity schemes declared in the Agent Card: API key, HTTP auth, OAuth 2.0, OpenID Connect, mutual TLSOptional OAuth 2.1 based authorisation for HTTP; stdio servers take credentials from the environment
Integrity of metadataAgent Cards may be signed with JWS (RFC 7515) after JCS canonicalisation (RFC 8785)Not defined for server metadata in the core specification
Current versionv1.0.1 (26 May 2026)Specification revision 2026-07-28
GovernanceAgentic AI Foundation project; Technical Steering CommitteeAgentic AI Foundation project; maintainers and the SEP process
LicenceApache License 2.0Transitioning from MIT to Apache-2.0; documentation under CC-BY-4.0

Recent milestones

As of 15 November 2026, dated from each project's changelog and announcements.

DateProtocolMilestone
9 December 2025MCPAnthropic donates MCP to the Agentic AI Foundation; goose and AGENTS.md are the other founding projects
12 March 2026A2Av1.0.0: a2a.proto becomes the normative source, enums move to SCREAMING_SNAKE_CASE, ListTasks added, implicit and password OAuth flows removed in favour of device code and PKCE, Agent Card signing specified
26 May 2026A2Av1.0.1: specification fixes, including preferring application/a2a+json in the HTTP binding
2026-07-28 (revision identifier)MCPSpecification revision 2026-07-28: stateless protocol, server/discover, sessions removed, Tasks moved to an extension, Roots, Sampling and Logging deprecated
17 August 2026A2AJoins the Agentic AI Foundation (AAIF announcement); an A2A project post of 27 August 2026 describes it as a Growth Stage project

Teams that adopted either protocol in 2025 should expect breaking changes on upgrade: A2A v1.0 lists its breaking changes in the changelog, and the MCP revision changes the request lifecycle itself.

How A2A works in brief

An A2A client fetches a remote agent's Agent Card to learn its skills, endpoints, supported bindings and security schemes, then authenticates and sends a message. The agent either answers directly with a Message or creates a Task that moves through lifecycle states until it is completed, failed, cancelled or rejected. The specification defines eleven core operations, mapped identically across bindings:

OperationJSON-RPC methodREST endpoint
Send messageSendMessagePOST /message:send
Send streaming messageSendStreamingMessagePOST /message:stream
Get taskGetTaskGET /tasks/{id}
List tasksListTasksGET /tasks
Cancel taskCancelTaskPOST /tasks/{id}:cancel
Subscribe to taskSubscribeToTaskPOST /tasks/{id}:subscribe
Push notification configs (create, get, list, delete)CreateTaskPushNotificationConfig and related methods/tasks/{id}/pushNotificationConfigs
Get extended Agent CardGetExtendedAgentCardGET /extendedAgentCard

Two features matter for security reviews. In-task authorisation lets an agent move a task to TASK_STATE_AUTH_REQUIRED so the client obtains a credential or human approval. The extended Agent Card lets an agent show more detail to authenticated clients than to anonymous ones. Both are examined in the A2A Protocol Security Guide.

How MCP works in brief

A host starts one MCP client per server. Under the 2026-07-28 specification there is no initialise handshake: each request carries its protocol version and client capabilities in _meta, and clients may call server/discover first to learn what a server supports. The model then calls tools, reads resources or uses prompts. When a server needs more input, it returns an input_required result under the Multi Round-Trip Requests pattern and the client retries with the answers. The same revision deprecated Roots, Sampling and Logging and reclassified the HTTP+SSE transport as Deprecated. The primitives are compared in MCP Tools vs Resources vs Prompts, and client support varies, as the MCP Client Support Tracker shows.

How they fit together

The A2A project's own comparison uses an auto repair shop. A customer, or the customer's assistant agent, talks to a Shop Manager agent over A2A. The manager assigns the job to a Mechanic agent, again over A2A. The mechanic uses MCP to call its tools: a diagnostic scanner, a repair manual database, a platform lift. When a part is needed, the mechanic contacts a Parts Supplier agent over A2A. A2A carries the conversational, task-oriented exchanges between agents and organisations; MCP carries the structured tool calls inside each agent.

  1. Inside one application: one or more agents use MCP servers for tools and data. A2A is not needed.
  2. Across teams or organisations: each agent keeps its own MCP tools private and publishes an Agent Card. Other agents delegate tasks over A2A without seeing those tools.
  3. Agent as a tool: the A2A documentation notes that a remote agent can expose well-defined, stateless skills in a tool-like way that an MCP-style description can represent. It also stresses that A2A's strength is stateful collaboration beyond a single tool call.

A gateway can sit in front of either protocol, and the two kinds of gateway solve different problems; see Agent Gateway vs MCP Gateway.

When to use which

Factual criteria for choosing a pattern, not a ranking:

SituationPattern that matches the protocol's stated purpose
An agent needs to query a database, call an API or read filesMCP server exposing tools or resources
A task must be delegated to an agent owned by another team or partner, whose internals you cannot seeA2A, with the remote agent's Agent Card as the contract
Work runs for hours and needs progress updates or human input part-wayA2A tasks with streaming or push notifications; in MCP, the opt-in Tasks extension where client and server support it
A coding assistant or desktop app needs local capabilitiesMCP over stdio
You need to prove that discovery metadata came from the claimed providerA2A signed Agent Cards; for MCP, rely on registry and transport controls
A simple request to a known HTTP service with no agent behaviourNeither may be needed; see A2A vs Direct API Integration and MCP vs Native Function Calling

Small teams building one product usually start with MCP alone. Larger organisations with many agents across business units, or agents that cross company boundaries, are where A2A's discovery, task model and declared security schemes become relevant.

Identity and security across both protocols

Each protocol adds machine-to-machine trust relationships, and they are different in kind:

  • MCP: the client holds a credential for each remote server (an OAuth token or API key), and a local stdio server runs with the user's privileges. Clients must validate a present iss parameter under RFC 9207 in the 2026-07-28 revision, and Client ID Metadata Documents are now preferred over Dynamic Client Registration. See the MCP Authorisation Specification Guide and the MCP Security Guide.
  • A2A: the calling agent authenticates to the remote agent using a scheme declared in its Agent Card, and server identity is verified at the transport layer. Push notifications introduce webhooks that need their own authentication. See the A2A Protocol Security Guide and the Multi-Agent and A2A Security Guide.
  • Delegation chains: when an A2A task leads to MCP tool calls, the user's original intent passes through several identities. Record which agent identity made each call and whether it acted for a user. The Agent Identity and Access Guide covers this, and Agent Cards vs OAuth Client Metadata compares the two discovery documents.

Common misconceptions

  • "A2A replaces MCP": the A2A project describes the two as complementary and recommends using both, A2A between agents and MCP for tools.
  • "MCP is only for local tools": MCP supports remote servers over Streamable HTTP with OAuth 2.1 based authorisation.
  • "An Agent Card is trusted because it is at a well-known URL": signing is optional (MAY) in the specification, so a client has to decide whether to require and verify signatures.
  • "MCP sessions keep state between calls": protocol-level sessions and the Mcp-Session-Id header were removed in 2026-07-28.
  • "Both protocols are owned by the companies that started them": both are Agentic AI Foundation projects; MCP maintainers and the A2A Technical Steering Committee set their technical direction.

Practitioner checklist

  • Map each integration as agent to tool (MCP) or agent to agent (A2A) before choosing a protocol, and document the decision.
  • Record the protocol version in use: A2A v1.0.x or earlier, and the MCP specification revision each client and server supports.
  • For A2A, decide whether you require signed Agent Cards and verify JWS signatures before trusting card contents.
  • For MCP, move remote servers to Streamable HTTP and plan for deprecated features (Roots, Sampling, Logging, HTTP+SSE).
  • Inventory every credential each protocol introduces: MCP client tokens and keys, A2A client credentials and push notification webhook secrets.
  • Log the full delegation chain across A2A tasks and the MCP tool calls they trigger, with the acting identity at each step.
  • Apply the same allowlist discipline to remote A2A agents as to MCP servers: only approved Agent Card URLs and server endpoints.

Standards and references

Related NHI Mgmt Group resources: MCP Servers Hub, MCP Specification Version Tracker, AI Agent Platforms Compared, Agent Skills vs MCP.

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 15 November 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security. nhimg.org

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick the PDF box to continue.