MCP vs A2A: Which Agent Protocol Do You Need in 2026?
MCP connects agents to tools and A2A connects agents to other agents. Compare scope, transport, adoption and payloads, with a decision table for real projects.
MCP (Model Context Protocol) connects one agent to tools, APIs and data sources. A2A (Agent2Agent) connects independent agents to each other so they can discover peers, delegate tasks and return results. They are not rivals. Most production systems that need both use A2A between agents and MCP inside each agent.
What is the difference between MCP and A2A?
The short version comes straight from the A2A specification. Its Appendix B says MCP focuses on "how AI models and agents connect to and interact with tools, APIs, data sources, and other external resources," while A2A focuses on "how independent, often opaque, AI agents communicate and collaborate with each other as peers."
The Linux Foundation, which hosts both projects, uses the same framing. Its April 9, 2026 press release calls A2A "complementary to the Model Context Protocol" and says A2A covers how agents coordinate "across organizational boundaries, while MCP defines how agents connect to internal tools and data sources."
A useful mental model: an MCP server is a capability (our TOON MCP server, for example, only converts JSON to TOON and back). It does what it is told and has no goals of its own. An A2A server is a collaborator. It receives a task, may plan and call its own tools, may ask for more input, and eventually returns artifacts.
MCP vs A2A: side-by-side comparison
This table uses the MCP 2026-07-28 specification and A2A v1.0.0, the current releases as of October 2026.
| Dimension | MCP (2026-07-28) | A2A (v1.0.0) |
|---|---|---|
| Connects | Agent or LLM app to tools, resources, prompts | Agent to agent, often across vendors or orgs |
| Core unit | Tool call with input and result | Task with lifecycle states, messages and artifacts |
| Wire format | JSON-RPC 2.0 (Streamable HTTP for remote servers) | JSON-RPC 2.0, gRPC or HTTP+JSON/REST bindings |
| State | Stateless core, no initialize handshake or session ID | Stateful tasks grouped by contextId |
| Discovery | tools/list with ttlMs and cacheScope cache hints | Agent Card at /.well-known/agent-card.json, optionally signed |
| Long-running work | Tasks extension (io.modelcontextprotocol/tasks) | Built in: polling, SSE streaming, push notifications |
| Payload content | Text, image, audio, resources, structuredContent | Parts: text, raw, url or data, each with optional mediaType |
| Typical caller | The model, through a host app | Another agent acting as an A2A client |
Sources: MCP 2026-07-28 release post, MCP tools spec, and the A2A specification.
How mature is each protocol in 2026?
Both are past the experimental stage. MCP shipped what its maintainers call a stateless protocol core on July 28, 2026. The release post says it "retired the initialize/initialized exchange along with the Mcp-Session-Id header," and that MCP's Tier 1 SDKs see "close to half-a-billion downloads a month." The TypeScript SDK alone grew from about 50.2M monthly downloads in January 2026 to about 219.4M in September, per npm download stats.
A2A reached v1.0, which the Linux Foundation describes as the first stable specification, with "Signed Agent Cards for cryptographic identity verification." The same release reports more than 150 supporting organizations, more than 22,000 GitHub stars on the core repository, and production SDKs in five languages (JavaScript, Java, Go, .NET and Python). Microsoft integrated A2A into Azure AI Foundry and Copilot Studio, and AWS added it to Amazon Bedrock AgentCore Runtime.
The scale is different, though. MCP is something nearly every agent framework speaks. A2A is the newer layer, adopted mainly where several independently built agents need to work together.
When do you need MCP, A2A, or both?
Pick by the boundary you are crossing, not by popularity.
| Your situation | Use | Why |
|---|---|---|
| One agent needs a database, GitHub, or an internal API | MCP | Deterministic capabilities, called by the model |
| Exposing your product to AI assistants | MCP | Hosts like IDEs and chat apps already speak it |
| A planner hands work to specialist agents you also own | Often neither; in-process calls or MCP | A protocol boundary adds latency and auth you may not need |
| Your agent delegates to another team's or vendor's agent | A2A | Opaque peers, discovery, task lifecycle, signed identity |
| Tasks take minutes or need human input midway | A2A (or the MCP Tasks extension) | Input-required states, streaming and push notifications |
| Cross-org agent network where each agent uses tools | Both | A2A between agents, MCP inside each one |
The third row is the one teams get wrong most often. If every agent runs in your own process and shares your auth model, wrapping each one in A2A mostly adds network hops. A2A earns its keep when the other side is opaque: you don't control its code, model or tools, and you need a contract.
What do the payloads look like?
Both protocols are JSON on the wire, but both also let you choose what the model actually reads. In MCP, a tool result carries structuredContent for programs and a text block for the model. In A2A v1.0, an artifact is a list of Parts, and the spec says a Part "MUST contain exactly one of the following: text, raw, url, data", with an optional mediaType on every part type.
That gives you a clean pattern for agent results that will be fed to another model. Put the machine copy in a data part and a token-efficient copy in a text part. Here is a flight-search agent returning three options. The TOON string below is real output from @toon-format/toon 4.1.1:
{
"artifactId": "flight-options",
"parts": [
{
"text": "flights[3]{id,from,to,depart,stops,price}:\n LH401,FRA,JFK,\"2026-11-03T10:15\",0,612\n UA961,FRA,EWR,\"2026-11-03T12:40\",0,588\n DL87,FRA,JFK,\"2026-11-03T13:05\",1,541",
"mediaType": "text/toon"
},
{
"data": {
"flights": [
{ "id": "LH401", "from": "FRA", "to": "JFK", "depart": "2026-11-03T10:15", "stops": 0, "price": 612 },
{ "id": "UA961", "from": "FRA", "to": "EWR", "depart": "2026-11-03T12:40", "stops": 0, "price": 588 },
{ "id": "DL87", "from": "FRA", "to": "JFK", "depart": "2026-11-03T13:05", "stops": 1, "price": 541 }
]
},
"mediaType": "application/json"
}
]
}Decoded, the text part is a plain TOON table:
flights[3]{id,from,to,depart,stops,price}:
LH401,FRA,JFK,"2026-11-03T10:15",0,612
UA961,FRA,EWR,"2026-11-03T12:40",0,588
DL87,FRA,JFK,"2026-11-03T13:05",1,541text/toon is the provisional media type in the TOON 4.1 spec, not a registered IANA type, so treat it as a label your own agents agree on. A receiving agent that doesn't understand it can fall back to the data part. On uniform arrays like this, TOON's official benchmark measured 42.6% fewer tokens than pretty JSON overall, and 58.7% fewer on flat tables like this one, with 72.2% vs 71.4% accuracy (effectively tied) across 5,856 calls. The same pattern on the MCP side is covered in MCP structuredContent vs text content.
Where does data format matter most in agent protocols?
Agent-to-agent hops multiply context. A planner that delegates to four specialists and reads four result sets pays for every token of those results, and then pays again on every later turn that keeps them in history. That is where a compact format helps most: large, uniform result sets that a model has to read.
Be careful about the other direction. Independent research is mixed for agents. The "Notation Matters" study (arXiv 2605.29676) tested token-optimized formats on tool-calling benchmarks with open-weight models and found TOON saved up to 18% of tokens at about a 9 percentage-point accuracy cost, concluding it "is not safe as a default in multi-turn agentic systems." We break that study down in TOON in multi-turn agents.
The practical rule: keep protocol envelopes, Agent Cards, tool inputs and anything a program parses as JSON. Convert only the large, read-only data that lands in a model's context, and measure on your own tasks before rolling it out. For a broader architecture view, see TOON for multi-agent systems.
Does MCP's 2026 spec change the comparison?
It narrows one gap and widens another. The stateless core means "any request can land on any instance behind a plain round-robin load balancer," according to the release post, and new Mcp-Method and Mcp-Name headers let gateways route without parsing bodies. That makes MCP servers easier to run at scale, which is a concern A2A's "web-aligned architecture" also targets.
At the same time, MCP moved Tasks out of the core and into an extension, and deprecated server-initiated features like Sampling and Roots (they keep working for at least twelve months). MCP is getting simpler and more tool-shaped. Long-running, multi-step collaboration between agents is squarely A2A's job. If you are tracking the MCP changes in detail, start with our MCP 2026-07-28 breakdown.
Frequently Asked Questions
Is A2A a replacement for MCP?
No. The A2A specification describes the two as complementary. MCP standardizes how an agent connects to tools, APIs and data sources. A2A standardizes how independent agents discover each other, delegate tasks and exchange results. A typical system uses A2A between agents and MCP inside each agent.
Do MCP and A2A both use JSON-RPC?
MCP messages are JSON-RPC 2.0. A2A v1.0 defines three protocol bindings: JSON-RPC 2.0, gRPC and HTTP+JSON/REST, and it uses Server-Sent Events for streaming. Both are JSON-first on the wire, so payload size and format matter in both.
How does an agent find another agent with A2A?
Through an Agent Card, a JSON document that describes the agent's identity, skills, endpoints and authentication. The spec's well-known location is /.well-known/agent-card.json on the agent's domain. Clients can also use registries or pre-configured card URLs, and v1.0 supports signed Agent Cards.
Can I send TOON over MCP or A2A?
Yes, as text. In MCP, put TOON in the text content block that the model reads and keep JSON in structuredContent. In A2A, a Part can carry text with a mediaType, so a text part can hold TOON while a data part holds the JSON for programmatic consumers.
Recommended Reading
MCP structuredContent vs Text Content: Where TOON Fits
MCP tools return structuredContent JSON plus a text block the model reads. Keep JSON for clients and put TOON in the text block. Spec rules, examples, caveats.
MCP 2026-07-28 Spec: Stateless Servers and Your Data
The MCP 2026-07-28 spec drops sessions and the initialize handshake, adds cacheable lists and header routing. What changed and how to format tool results.
Is TOON Safe for AI Agents? What the 2026 Research Says
A 2026 arXiv study found TOON saves up to 18% of tokens in agent loops but costs about 9 accuracy points on open models. Where it helps and where it breaks.