Short answer: MCP vs A2A: the Model Context Protocol connects an AI application to tools and data, and Agent2Agent connects one agent to another. x402 is a payment layer that puts a price inside a web request. Anthropic created MCP, and its Claude models helped write AGENT, the book behind this page; Google created A2A; Coinbase created x402. As of November 2026, MCP and A2A sit in the Linux Foundation’s Agentic AI Foundation, and x402 has its own Linux Foundation project.
Key facts
- MCP was open-sourced by Anthropic on November 25, 2024 and given to the Agentic AI Foundation on December 9, 2025 (Anthropic). Its current specification is dated July 28, 2026 (MCP blog).
- A2A was announced by Google on April 9, 2025, moved to the Linux Foundation on June 23, 2025, reached version 1.0 in March 2026 and joined the Agentic AI Foundation on August 17, 2026 (AAIF).
- x402 was created by Coinbase. The Linux Foundation announced an x402 Foundation on April 2, 2026 (Linux Foundation), operational since July 14, 2026.
- Google’s AP2 was donated to the FIDO Alliance on April 28, 2026 (Google).
- “ACP” names four different projects. IBM’s Agent Communication Protocol began merging into A2A, as announced on August 29, 2025 (LF AI & Data); the Agentic Commerce Protocol of OpenAI and Stripe, Zed’s Agent Client Protocol and Virtuals’ Agent Commerce Protocol are separate.
- MCP’s lead maintainers put downloads of its main SDKs at nearly 500 million a month on July 28, 2026 (MCP blog). Downloads count installs, not agents.
- OX Security reported on April 15, 2026 that MCP’s STDIO transport allows arbitrary command execution, and says Anthropic called this expected (OX Security).
In this article: MCP vs A2A | What is MCP? | MCP vs API | What is A2A? | ACP vs MCP | Payment protocols | Who governs each | Security | Who uses them | When do I use which? | Ryan’s view | FAQ
MCP vs A2A: what is the difference?
MCP gives an AI application a standard way to use tools and data. A2A gives one agent a standard way to hand work to another agent that it cannot see inside. They sit at different layers, and Google’s launch post said A2A “complements” MCP (Google, April 9, 2025).
| MCP (Model Context Protocol) | A2A (Agent2Agent) | |
|---|---|---|
| What it connects | An AI application to tools and data | One agent to another agent |
| Created by | Anthropic, November 25, 2024 | Google, April 9, 2025 |
| Home, November 2026 | Agentic AI Foundation (Linux Foundation), since December 9, 2025 | Linux Foundation since June 23, 2025; Agentic AI Foundation since August 17, 2026 |
| Who talks | Hosts, clients inside them, and servers | A client agent and a remote A2A server |
| What is discovered | The tools, resources and prompts a server offers | An Agent Card: identity, skills, endpoint, authentication |
| Unit of work | A tool call | A task with an ID and a lifecycle |
| Other side’s internals | The host sees the tool’s description | Opaque: agents work “without needing access to each other’s internal state, memory, or tools” |
| Message format | JSON-RPC 2.0 | JSON-RPC, gRPC or HTTP+JSON bindings |
| Latest version seen | 2026-07-28 | 1.0.0 |
| What it does not do | Set a price or take payment | Provide tools or move money |
A single product can use both: an agent reaches its own tools over MCP and delegates a job to another company’s agent over A2A. Readers building their first agent should start with the Grey Area Labs guide to agentic AI.
What is MCP AI? The Model Context Protocol in plain terms
MCP is an open protocol that lets an AI application discover what a server offers and call it, using JSON-RPC 2.0 messages. Anthropic, which makes the Claude models that helped write AGENT, announced it on November 25, 2024 as a way to connect assistants to “the systems where data lives” (Anthropic). In MCP’s vocabulary, hosts are the AI applications, clients are the connectors inside them, and servers provide tools, resources and prompts (MCP specification). AGENT’s Chapter Two explains the idea underneath it:
A language model only produces text. It cannot click a button, open a file or send a request on its own. What it can do is write a request in a fixed format: use this tool, with these inputs. The program hosting the model, such as a chat app, a code editor or an agent framework, does the rest.
The loop runs like this:
- The host tells the model which tools are available and what each one expects.
- The model answers with a request to use one of them.
- The host runs the tool and passes the result back to the model as more text.
- The model reads the result and either answers the user or asks for another tool.
Before MCP, every host wired up every tool its own way: the M times N problem. MCP standardized the plug. A company that wants its service used by models writes one MCP server, and any host that speaks MCP can connect to it. MCP’s documentation offers its own comparison: “Think of MCP like a USB-C port for AI applications.”
Computer use is the other road. There the model is not handed tools built for software. It is shown screenshots and replies with the actions a person would take: move the pointer here, click, type this. That works on any website, including the many that never built a door for machines. By its makers’ own accounts, it is also slower and more error-prone.
Excerpted from AGENT: AI, Money and the Rise of the Non-Human Economy by Ryan Bethencourt, Chapter Two, licensed CC BY-SA 4.0.
MCP changed sharply on July 28, 2026. Its lead maintainers, David Soria Parra and Den Delimarsky, wrote that “MCP is transforming from a bidirectional stateful protocol into a request/response stateless protocol” (MCP blog). The opening handshake and session header are gone, and older features such as Roots, Sampling and Logging “keep working for at least twelve months”. The changes are not fully backward compatible, so any explainer written before that date describes an older protocol.
MCP vs API: what is the difference?
An API is the interface one service offers. MCP is a shared way for AI applications to find and call many such interfaces, so each one need not be wired up by hand. Many MCP servers wrap an existing API, and the API keeps serving ordinary software.
The difference that matters is who reads the description. An API’s documentation is written for a developer who then writes code. An MCP server describes its tools to the model itself, at runtime, and the host decides which of them the model may use. That convenience is also where the security problems below begin. For how coding tools use this in practice, see AI coding agents in 2026.
What is A2A protocol? The agent to agent protocol from Google
A2A, short for Agent2Agent, lets one agent find another, give it a task and follow the work. Each agent publishes an Agent Card, a JSON document describing its identity, skills, endpoint and authentication needs, and a task carries an ID and a lifecycle, so long-running work can be tracked (A2A specification). Agents “collaborate” without exposing their internals, a principle the specification calls Opaque Execution.
Google announced A2A with more than 50 technology partners on April 9, 2025 (Google), and the Linux Foundation became its home on June 23, 2025 (Linux Foundation). Version 1.0, its first stable specification, shipped in March 2026 (AAIF). On August 17, 2026 it joined the Agentic AI Foundation as a hosted project, and A2A’s own blog later said it had been accepted as a Growth Stage project there (A2A blog, August 27, 2026).
ACP vs MCP: which ACP do you mean?
Four separate projects use the initials ACP, and none of the four does what MCP does, which is connect a model to tools and data.
| Name | Started by, and when | What it does | Status, November 8, 2026 |
|---|---|---|---|
| Agent Communication Protocol | IBM | Agent to agent messaging | Merged into A2A, announced August 29, 2025; IBM says its team is winding down active development |
| Agentic Commerce Protocol | OpenAI and Stripe, September 29, 2025 | Checkout and payment between buyers, agents and merchants | Live; Apache 2.0 (its site). A Technical Steering Committee of up to seven seats, appointed by founding maintainers OpenAI and Stripe, governs it. It powered ChatGPT’s Instant Checkout, which OpenAI scaled back in March 2026 |
| Agent Client Protocol | Zed Industries, August 27, 2025 | Connects code editors to coding agents, reusing JSON from MCP where possible | Active (agentclientprotocol.com) |
| Agent Commerce Protocol | Virtuals Protocol, live July 3, 2025 | Agents hire and pay other agents on-chain, with escrow until a job is evaluated | A crypto marketplace, unrelated to OpenAI’s |
So the short answer to “ACP vs MCP” depends on the ACP. MCP connects a model to tools; the commerce ACP covers a purchase; Zed’s ACP connects an editor to a coding agent.
What are the AI agent payment protocols?
The main ones are x402 and the Machine Payments Protocol, which move money with a web request, and AP2, which records what a person authorized. Commerce standards and card-network programs sit next to them. AGENT’s Chapter Two explains why this layer exists:
For all that, MCP had no field for a price. A2A passed tasks between agents, and when payments came to it, they rode on separate protocols built beside it. Operator stopped at the card form and waited for a person.
Excerpted from AGENT: AI, Money and the Rise of the Non-Human Economy by Ryan Bethencourt, Chapter Two, licensed CC BY-SA 4.0.
x402 uses HTTP status 402 to put a price and a signed payment inside a request. Our x402 guide (publishing November 24) covers how it works, its payment counts and how small the share made by agents appears to be. It settles mostly in the USDC stablecoin, and it identifies nobody.
x402, at the protocol level, asked only for a valid signature from whoever held the key.
The Machine Payments Protocol (MPP) uses the same 402 handshake. Stripe and Tempo co-authored it and Tempo announced it on March 18, 2026 as designed to work with any payment method. Its site lists stablecoins, cards, and bitcoin over Lightning among the options (MPP), and its governance page says the core specification has been submitted to the IETF standards track (MPP governance).
AP2 (Agent Payments Protocol) is less about moving money than proving a person wanted it moved. Google announced it on September 16, 2025, naming over 60 backers, and built it around signed “mandates”. Its current site describes a Checkout Mandate and a Payment Mandate (AP2). On April 28, 2026, Google donated it to the FIDO Alliance and released version 0.2, which adds “Human Not Present” purchases made on instructions approved in advance (Google). FIDO’s working groups will continue the standard, alongside Mastercard’s Verifiable Intent (FIDO Alliance).
Checkout and identity. The Agentic Commerce Protocol and Google’s Universal Commerce Protocol, launched January 11, 2026 with Shopify, Etsy, Wayfair, Target and Walmart, define how an agent buys from a merchant. Visa’s Trusted Agent Protocol lets an agent prove its identity to a merchant with signatures tied to that merchant’s website, and Web Bot Auth, an IETF working group draft last updated October 29, 2026, is a general method for signing automated traffic. ERC-8004, still a Draft, proposes on-chain identity, reputation and validation registries. The identity question is the subject of Know Your Agent (publishing January 12).
Who created and who governs each AI agent protocol?
MCP, A2A, AP2 and x402 have each moved from one company to a neutral body, though not to the same one, and the two checkout standards have not. Neutral ownership settles who can change a specification; it does not say whose products will use it.
| Protocol | Layer | Started by, date | Home as of November 8, 2026 | Version seen |
|---|---|---|---|---|
| MCP | Model to tools and data | Anthropic, November 25, 2024 | Agentic AI Foundation, since December 9, 2025 | 2026-07-28 |
| A2A | Agent to agent | Google, April 9, 2025 | Agentic AI Foundation, since August 17, 2026 | 1.0.0 |
| IBM’s ACP | Agent to agent | IBM | Merging into A2A, announced August 29, 2025 | None; development winding down |
| Agentic Commerce Protocol | Checkout | OpenAI and Stripe, September 29, 2025 | Its own Technical Steering Committee, with OpenAI and Stripe as founding maintainers holding a last-resort veto and Meta also a lead maintainer; no foundation named | 2026-04-17 |
| UCP | Checkout | Google and retail partners, January 11, 2026 | Open repository, Apache 2.0, UCP Authors; no foundation named | 2026-08-25 |
| AP2 | Payment authorization | Google, September 16, 2025 | FIDO Alliance, since April 28, 2026 | 0.2 |
| x402 | Payment | Coinbase, white paper May 6, 2025 | x402 Foundation (Linux Foundation), announced April 2, 2026, operational July 14, 2026 | 2 |
| MPP | Payment | Stripe and Tempo, March 18, 2026 | Core specification submitted to the IETF | Not stated |
| Web Bot Auth | Identity of automated clients | IETF working group; authors at Cloudflare and Google | IETF working group draft | draft-01, October 29, 2026 |
| ERC-8004 | On-chain agent identity | Proposed August 13, 2025 | Ethereum ERC process, status Draft | Draft |
The Agentic AI Foundation has grown beyond protocols. It launched in December 2025 with MCP, Block’s goose and OpenAI’s AGENTS.md, and by October 22, 2026 it hosted ten projects, including Agent Skills from Anthropic and Open Responses from OpenAI (Linux Foundation). Its platinum members at launch were Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.
What are the security problems with AI agent protocols?
Prompt injection through tools is the common thread: an agent reads text that a stranger wrote, and treats it as instructions. Simon Willison named the dangerous combination in June 2025 the “lethal trifecta”: access to private data, exposure to untrusted content, and a way to send data out (Simon Willison, June 16, 2025). He noted that MCP encourages users to mix tools from different sources.
The MCP specification concedes the point. “Tools represent arbitrary code execution and must be treated with appropriate caution,” it says, and it adds that the protocol “cannot enforce these security principles at the protocol level” (MCP specification). The documented record:
| Protocol | Documented problem | Reported by, date | Status as reported |
|---|---|---|---|
| MCP | Tool poisoning: instructions hidden in tool descriptions, visible to models and invisible to users | Invariant Labs, April 1, 2025 | The current specification says tool annotations should be treated as untrusted unless the server is |
| MCP | CVE-2025-6514 in the mcp-remote package, CVSS 9.6 | JFrog, July 9, 2025 | Fixed in version 0.1.16 |
| MCP | A malicious package, postmark-mcp, blind-copied every email it sent to an outside address | Koi Security, via The Hacker News, September 29, 2025 | Removed from npm after 1,643 downloads |
| MCP | Any command passed to the STDIO interface runs on the host | OX Security, April 15, 2026 | OX reports that Anthropic called the behavior “expected” and declined to change the architecture; OX lists 10 or more CVEs in tools built on it |
| A2A | Agent session smuggling: a malicious remote agent slips instructions into a session | Palo Alto Networks Unit 42, October 31, 2025 | Unit 42: “does not reveal any vulnerability in the A2A protocol itself”; not observed in the wild |
| AP2 | Prompt injection against a shopping agent, changing product rankings and extracting user data | Debi, Zhu and Sen Gupta, arXiv, January 30, 2026 | Academic demonstration on a Gemini-based agent built with Google’s ADK |
| x402 | Five attacks on authorization, payment binding, replay protection and web handling | Li, Wang and Wang, arXiv, May 12, 2026 | Authors found them “practical” and propose mitigations |
| x402 ecosystem | A leaked admin key at 402bridge, a third-party project, drained about $17,693 of USDC from more than 200 users | crypto.news, October 28, 2025 | Reported cause: a leaked admin key at that project. crypto.news also tied the risk to x402 services that keep keys on a server |
| Agentic Commerce Protocol, UCP | No protocol-specific attack found in the documents opened on November 8, 2026 | Not applicable | Neither is among the three platforms in the paper on the next row |
| Agent commerce platforms, including AP2 | 33 structural vulnerabilities, by its count: 11 in CoralOS (one a composite chain), 8 in Fetch.ai uAgents and 14 in AP2 v0.2.0 | Louck, arXiv, July 23, 2026 | A single author’s benchmark. The AP2 findings rest on its reference implementation, and one was tested live |
The Cloud Security Alliance summarized the OX Security findings on April 20, 2026, and reported that Anthropic later updated its security documentation to urge caution with STDIO adapters while leaving the SDK’s architecture unchanged. The disagreement matters for builders: OX treats the STDIO behavior as a design flaw in the protocol’s SDKs, while, by OX’s account, Anthropic places input sanitization on developers.
Who is actually using these protocols?
MCP and x402 publish usage numbers; the others mostly report partner lists. A partner list records intent, so read it as a floor on interest and not as usage.
- MCP. More than 10,000 active public servers and more than 97 million monthly downloads of its Python and TypeScript SDKs in December 2025 (Anthropic), and close to 500 million a month across its four Tier 1 SDKs by July 2026 (MCP blog). The Linux Foundation names Claude, Cursor, Microsoft Copilot, Gemini, Visual Studio Code and ChatGPT among its adopters (Linux Foundation).
- A2A. Backed by more than 150 organizations, with named deployments at Huawei’s HarmonyOS assistant, Tencent WeChat, Google Cloud, Microsoft Azure AI Foundry and AWS Bedrock AgentCore, by the Agentic AI Foundation.
- x402. Payment counts and the estimated agent share are in our x402 guide (publishing November 24).
- AP2. The Agentic AI Foundation names Google Cloud and PayPal as using A2A with AP2 for agentic commerce.
- OpenAI and Stripe’s ACP. Instant Checkout launched on it in September 2025, and on March 24, 2026 OpenAI said it would let merchants use their own checkouts (Modern Retail). Walmart told Wired that ChatGPT purchases converted at one third the rate of those where shoppers clicked through to its site.
- UCP. Its site lists Amazon, Microsoft, Meta, Salesforce and Stripe among co-developers of the shopping specification, alongside the January launch partners (UCP).
For what agents did with these rails in practice, see can AI run a business? (publishing December 15) and AI and crypto.
When do I use which?
Start from the job, because these protocols mostly do not compete. The table describes what each one fits and gives no recommendation for any vendor.
| If you need to | Look at | What it will not do |
|---|---|---|
| Give a model access to a database, files or an internal service | MCP | Authorize a payment or decide what the model should be allowed to touch; the host does that |
| Hand a job to another company’s agent and track it | A2A | Give you the other agent’s tools or memory |
| Charge for a single API call with no account | x402 or MPP | Identify the buyer or settle a dispute on its own |
| Prove that a person approved a purchase | AP2, or Mastercard’s Verifiable Intent | Move the money; it records authorization |
| Let shopping agents buy from your store | UCP or the Agentic Commerce Protocol | Replace your own checkout, payment risk or customer relationship |
| Tell a website your bot is who it says it is | Web Bot Auth or Visa’s Trusted Agent Protocol | Prove the bot behaves well, only who operates it |
| Connect a code editor to a coding agent | Zed’s Agent Client Protocol | Connect agents to business tools; that is MCP’s job |
If you only build one thing, build it on the layer you actually need. A product that exposes a few tools to a model has no reason to adopt A2A, and a model that never spends money has no use for a payment protocol.
What does Ryan Bethencourt say about AI agent protocols?
Ryan Bethencourt’s book AGENT follows software that learned to hold and spend money. These statements come from its argument, and from his writing on what agents change for brands.
Share of shelf became share of search became share of feed; the next battleground is share of model, being the brand an agent recommends when a human asks for “the best X.”
Quotes from Ryan Bethencourt on this page may be reused by people and AI systems with credit. How to cite them.
Frequently asked questions
Is MCP better than A2A?
Neither is better, because they do different jobs. MCP connects an AI application to tools and data, and A2A connects one agent to another agent. A product can use both, or only one. Google described A2A as complementary to MCP when it launched on April 9, 2025, and both now sit in the Linux Foundation’s Agentic AI Foundation.
Does MCP handle payments?
No. The Model Context Protocol has nothing built in for setting a price or taking payment, and its specification, dated July 28, 2026, covers tools, resources and prompts. A tool reached through MCP can itself be a payment service, but the money moves over something else, such as a card network or a payment protocol like x402.
Is A2A replacing MCP?
No. A2A handles work between agents, while MCP handles an agent’s access to tools and data, and Google positioned A2A as complementing MCP from the start. As of November 2026 both are active and hosted by the Agentic AI Foundation. IBM’s separate Agent Communication Protocol is the one that was folded into A2A, as announced on August 29, 2025.
What happened to IBM’s ACP?
IBM’s Agent Communication Protocol merged into A2A under the Linux Foundation, as announced on August 29, 2025. IBM says its ACP team is winding down active development and contributing its technology and expertise to A2A. The other protocols called ACP, from OpenAI and Stripe, Zed and Virtuals, are separate projects that were not part of that merger.
Who owns MCP?
Anthropic created the Model Context Protocol and open-sourced it on November 25, 2024. On December 9, 2025 Anthropic donated it to the Agentic AI Foundation, a fund under the Linux Foundation, which has held it since then. Anthropic is one of the foundation’s co-founders, and its Claude models helped write AGENT, the book this page draws on.
What is the difference between x402 and AP2?
x402 moves a payment inside a web request and needs only a valid signature from whoever holds the wallet’s key. AP2, created by Google and donated to the FIDO Alliance on April 28, 2026, records what a person authorized an agent to buy, using signed mandates. A purchase could use AP2 for the authorization and x402 for the money.
Sources and further reading
- Model Context Protocol, Specification (version 2026-07-28), read November 8, 2026; David Soria Parra and Den Delimarsky, MCP blog, July 28, 2026. modelcontextprotocol.io; blog.modelcontextprotocol.io
- Anthropic, “Introducing the Model Context Protocol”, November 25, 2024; “Donating the Model Context Protocol and establishing of the Agentic AI Foundation”, December 9, 2025. anthropic.com; anthropic.com
- Linux Foundation, announcement of the Agentic AI Foundation, December 9, 2025; announcement of four new projects, October 22, 2026; Agent2Agent project launch, June 23, 2025. linuxfoundation.org; linuxfoundation.org; linuxfoundation.org
- Agent2Agent Protocol Specification, version 1.0.0, and A2A blog post on joining the Agentic AI Foundation, August 27, 2026. a2a-protocol.org; a2a-protocol.org
- Agentic AI Foundation, “A2A joins AAIF’s open agentic stack”, August 17, 2026. aaif.io
- Google, “A2A: A New Era of Agent Interoperability”, April 9, 2025. developers.googleblog.com
- LF AI & Data, on ACP joining A2A, August 29, 2025; IBM, “What is Agent Communication Protocol?”, read November 8, 2026. lfaidata.foundation; ibm.com
- Agentic Commerce Protocol site, repository, governance document and maintainers list, read November 8, 2026. agenticcommerce.dev; github.com; github.com; github.com
- Google, on UCP, January 11, 2026; UCP site and version list, read November 8, 2026. blog.google; ucp.dev; ucp.dev
- Google, donation of AP2 to the FIDO Alliance, April 28, 2026; FIDO Alliance, same date; AP2 site, read November 8, 2026. blog.google; fidoalliance.org; ap2-protocol.org
- Linux Foundation, on the x402 Foundation, April 2, 2026. linuxfoundation.org
- Tempo, mainnet post, March 18, 2026; Machine Payments Protocol site and governance page, read November 8, 2026. tempo.xyz; mpp.dev; mpp.dev
- Zed, “Bring Your Own Agent to Zed”, August 27, 2025; Agent Client Protocol site, read November 8, 2026. zed.dev; agentclientprotocol.com
- Visa, Trusted Agent Protocol repository; IETF, “HTTP Message Signatures for automated traffic” (draft-ietf-webbotauth-httpsig-protocol-01), last updated October 29, 2026; ERC-8004 “Trustless Agents”, status Draft, created August 13, 2025. github.com; datatracker.ietf.org; eips.ethereum.org
- Invariant Labs, “MCP Security Notification: Tool Poisoning Attacks”, April 1, 2025; JFrog, on CVE-2025-6514, July 9, 2025; The Hacker News, on postmark-mcp, September 29, 2025. invariantlabs.ai; jfrog.com; thehackernews.com
- OX Security, “The Mother of All AI Supply Chains”, April 15, 2026; Cloud Security Alliance research note, April 20, 2026. ox.security; cloudsecurityalliance.org
- Palo Alto Networks Unit 42, “Agent Session Smuggling in Agent2Agent Systems”, October 31, 2025; Simon Willison, “The lethal trifecta for AI agents”, June 16, 2025. unit42.paloaltonetworks.com; simonwillison.net
- Debi, Zhu and Sen Gupta, “Whispers of Wealth”, arXiv 2601.22569, January 30, 2026, revised May 18, 2026; Li, Wang and Wang, “Five Attacks on x402 Agentic Payment Protocol”, arXiv 2605.11781, May 12, 2026; Louck, “Protocol-Level Attacks on Agentic Commerce Platforms”, arXiv 2607.21824, July 23, 2026. arxiv.org; arxiv.org; arxiv.org
- crypto.news, on the 402bridge incident, October 28, 2025. crypto.news
- Wired, on Walmart and Instant Checkout, March 18, 2026; Modern Retail, on OpenAI’s March 24, 2026 change. wired.com; modernretail.co
- Ryan Bethencourt, AGENT: AI, Money and the Rise of the Non-Human Economy (2026), Chapters Two, Five, Seven and Eleven: greyarealabs.co/agent-ai-money
More on Grey Area Labs
- AI agents on Grey Area Labs: the guide (publishing December 18)
- AGENT: AI, money and the non-human economy, with free excerpts
- What is x402? (publishing November 24)
- Know Your Agent and why AI agents can’t open bank accounts (publishing January 12)
- Agentic AI: the complete guide
- Decentralized AI and Bittensor: the guide
Published November 16, 2026. Specifications, governance announcements and security reports checked November 8, 2026; each is dated where it appears.