# Agent protocols in September 2025: MCP gets a registry, AP2 adds payments

- **Published:** 2025-09-30
- **Updated:** 2026-10-12
- **Category:** Harnesses & Tooling
- **Author:** Sinthetix
- **URL:** https://sinthetix.com/blog/agent-protocols-mcp-a2a-ap2

> The MCP Registry preview and Google's AP2 payments protocol show agent interoperability maturing in layers. What each does and where the gaps remain.

Agent interoperability is settling into layers. One standard connects models to tools and data. Another lets agents talk to each other. September added two more pieces: a public catalog for the first, and a payments protocol that sits alongside both. For anyone building or buying agent systems, the shape of those layers is more useful to understand than any single announcement.

**Key takeaways**

- MCP, which connects models to tools and data, was adopted by OpenAI in March; its June spec revision tightened authorization.
- The MCP Registry launched in preview on September 8 as a catalog of public servers, with a warning that breaking changes may occur.
- Google's AP2, announced September 16, defines signed "mandates" so an agent can pay on a user's behalf, with more than 60 collaborating organizations.
- All of these are young; the security and governance work sits with the adopter.

## The layers

| Layer | Protocol | What it does | Latest development |
| --- | --- | --- | --- |
| Model to tools and data | MCP | Standard way for apps to give models context and tools | Registry preview, Sep 8 [4] |
| Agent to agent | A2A | Lets agents find each other and work across vendors | Linux Foundation project, Jun 23 [2] |
| Agent to payment | AP2 | Authorization and audit trail for agent-led purchases | Announced Sep 16 [5] |

## MCP: adoption, then hardening

In March, OpenAI said it would support Anthropic's Model Context Protocol across its products; Sam Altman wrote, "People love MCP and we are excited to add support across our products" [1]. MCP support was already in OpenAI's Agents SDK, with ChatGPT desktop and the Responses API to follow [1]. Block, Apollo, Replit, Codeium and Sourcegraph were among others that had added it [1].

Wider adoption exposed security gaps, and the June 18 spec revision reflects that. It classifies MCP servers as OAuth resource servers, requires clients to implement Resource Indicators (RFC 8707) to prevent malicious servers from obtaining access tokens, adds a security best-practices page, introduces elicitation so servers can request more information from users, and adds structured tool output [3].

## The MCP Registry

On September 8, the project launched the MCP Registry in preview: an open catalog and API for publicly available MCP servers, aimed at discoverability [4]. Maintainers publish self-reported information, and the community can flag servers that violate moderation guidelines, which maintainers can denylist [4]. Organizations can build sub-registries on top, public ones such as marketplaces tied to clients, or private ones inside enterprises with strict security requirements [4]. Maintainers named include people from Anthropic, PulseMCP and GitHub, and the initial announcement listed 16 contributors from at least 9 companies [4].

Two caveats come from the announcement itself. The preview "does not provide data durability guarantees," and "breaking changes may occur before the registry is made generally available" [4]. And the data is self-reported [4]: a listing is a discovery aid, not a trust signal. The private sub-registry model is the one enterprises should note, since it allows an approved list of servers rather than open access.

## A2A: agents across vendors

The Agent2Agent protocol, created by Google, moved to the Linux Foundation on June 23, with more than 100 technology companies supporting it [2]. Its purpose is to let autonomous agents find each other, exchange information securely and cooperate across systems and vendors, under vendor-neutral governance [2].

## AP2: agents that pay

On September 16, Google announced the Agent Payments Protocol, an open protocol built with payments and technology companies for agent-led payments [5]. The problem it targets: existing payment systems assume a human clicking "buy" [5]. AP2 uses cryptographically signed mandates. An Intent Mandate captures the user's request, including rules such as price limits and timing for delegated tasks, and a Cart Mandate locks the exact items and price once approved, together forming an audit trail from intent to payment [5]. It supports cards, stablecoins and real-time bank transfers, and can extend A2A and work with MCP [5]. Google says it is collaborating with more than 60 organizations, including Adyen, American Express, Coinbase, Mastercard and PayPal [5].

AP2 is a specification with reference implementations at this stage [5]. The announcement describes a companion A2A x402 extension for crypto payments as "production-ready" [5], but it does not report deployments of AP2 itself, and neither do we.

## What is still unsettled

- **Trust in catalogs.** A registry of self-reported servers helps find tools, not vet them [4].
- **Authorization in practice.** The June spec sets the model, but implementations vary by client and server [3].
- **Accountability for money.** Mandates define evidence of intent, but operational responsibility for agent purchases is a matter for the surrounding systems and payment networks, not the protocol.

## Practical implication

Build agent integrations on open protocols where you can, but treat third-party MCP servers like any vendor software: vet, pin and approve them, ideally through a private registry. Do not allow agents to spend money until limits, approvals and audit trails are defined.

## Sources

1. [OpenAI adopts rival Anthropic's standard for connecting AI models to data](https://techcrunch.com/2025/03/26/openai-adopts-rival-anthropics-standard-for-connecting-ai-models-to-data/), TechCrunch, 2025-03-26
2. [Linux Foundation launches the Agent2Agent protocol project](https://www.linuxfoundation.org/press/linux-foundation-launches-the-agent2agent-protocol-project-to-enable-secure-intelligent-communication-between-ai-agents), Linux Foundation, 2025-06-23
3. [MCP specification changelog, version 2025-06-18](https://modelcontextprotocol.io/specification/2025-06-18/changelog), Model Context Protocol, 2025-06-18
4. [Introducing the MCP Registry (preview)](https://blog.modelcontextprotocol.io/posts/2025-09-08-mcp-registry-preview/), Model Context Protocol blog, 2025-09-08
5. [Announcing Agents to Payments (AP2) Protocol](https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol), Google Cloud, 2025-09-16

## How Sinthetix helps

**AI Agents & Automation**: Intelligent agents and automated workflows that reduce manual work and streamline operations.

- Multi-agent systems
- Task automation
- AI-driven decision logic

[Explore AI Agents & Automation](https://sinthetix.com/services/ai-agents-automation) | [Talk to us](https://sinthetix.com/contact)

Sinthetix is an AI engineering studio. We ship agents, automation, and internal tools for operators who need results, not reports. [About Sinthetix](https://sinthetix.com/about)