How AI agent marketplaces could become the app stores of a new agentic era

OpenAI, Anthropic, Google and Microsoft are building directories for AI agents and connected tools, setting up a new battle over software discovery, payments and platform control

OpenAI, Anthropic, Google and Microsoft are racing to make AI agents, rather than apps, the main way people use software. The smartphone era was built around app stores. The AI era could be shaped by "agent marketplaces": directories built into ChatGPT, Claude, Gemini and Copilot that allow an AI assistant to choose which outside tool or service to use on a person's behalf.

 

The early battles over who gets listed, who earns a share of transactions and who sets the rules resemble the beginning of the smartphone era, but with a different set of potential gatekeepers.

From AI chat to AI action

This shift can be easy to miss when viewed only through the chat window. On the surface, the experience remains similar: users type a request into a text box. What happens after they submit that request, however, is changing.

 

 

Through 2025 and into 2026, major AI companies pushed their chatbots beyond answering questions towards completing tasks, such as booking a table, buying a product or retrieving information from a company's internal systems.

 

To perform such tasks, an AI assistant needs a way to access and use outside software.

Much of this integration is increasingly based on the Model Context Protocol (MCP), an open standard introduced by Anthropic in November 2024. MCP is designed to connect AI assistants with systems containing data and tools, including content repositories, business software and development environments.

In simple terms, MCP acts as a common plug. Rather than building separate connections for every AI platform, a developer can create an MCP server that can be used by compatible AI applications.

OpenAI, Google and Microsoft have since added support for MCP alongside Anthropic. OpenAI's Apps Software Development Kit (SDK), for instance, requires developers to use an MCP server to define an application's capabilities and connect them with ChatGPT.

 

Anthropic said in late 2025 that Claude's directory had more than 75 MCP-powered connectors. Industry estimates cited by security research firm Practical DevSecOps suggest that more than 10,000 public MCP servers are available and monthly downloads of MCP-related toolkits have crossed 97 million.

 

A separate standard developed by Google, called Agent2Agent (A2A), performs a different role. While MCP helps an agent connect to tools and data, A2A is intended to let one AI agent communicate with another.

 

This matters commercially because once an AI assistant can reliably use outside services, the platform deciding which services are presented to users controls a part of the software-discovery business previously occupied by the app icon.

The battle to build agent marketplaces

Major AI companies have begun developing their own versions of an agent marketplace. Their approaches provide an early indication of how each company believes the emerging market should operate.

 

OpenAI

 

OpenAI has developed one of the most consumer-facing versions of the model. The company opened its Apps SDK to outside developers in 2025 and subsequently introduced an in-product directory in ChatGPT where approved apps could be discovered. Early apps included services from Spotify, Adobe, Canva, Booking.com, MyFitnessPal and AllTrails.

 

OpenAI has since reorganised the system around plugins. Its developer documentation says apps are now submitted and published as plugins, with applications, skills and other capabilities forming a broader discovery layer across ChatGPT and its coding product, Codex.

 

The payment infrastructure is also taking shape. Stripe-powered checkout allows users in supported markets to complete purchases without leaving the chat. Stripe issues a scoped payment token that lets ChatGPT initiate the transaction without exposing the buyer's underlying payment credentials.

 

Anthropic

 

Anthropic, which created MCP, has adopted a quieter and more developer-focused approach.

Its Connectors Directory lists tools that users can link to Claude. These include remote services such as Notion, Canva and Stripe, as well as applications such as Figma and Prisma.

The connectors work across Anthropic products, including Claude on the web and desktop, Claude Code and its application programming interface (API).

 

Unlike OpenAI's emerging commerce model, Anthropic does not yet offer an integrated marketplace payment system through which developers can sell connectors. That difference affects how developers building tools for Claude can earn revenue from them.

 

Google

 

At its Cloud Next 2026 conference, Google positioned Gemini Enterprise as an end-to-end platform for building, deploying and managing AI agents.

 

The Gemini Enterprise Agent Platform gives organisations a central place to build and govern agents, connect them with company data and make them available to employees. It also includes a registry through which organisations can store, discover and manage agents, tools and MCP servers.

Google's strategy is centred on providing the broader infrastructure around agents — from models and cloud computing to data connections, security and governance — rather than focusing only on a consumer-facing directory.

 

The platform can also support models from outside Google, including those developed by Anthropic, reinforcing its pitch as an open enterprise platform rather than a marketplace restricted to Gemini models.

 

Microsoft

 

Microsoft is building on its existing base of enterprise customers and Microsoft 365 subscribers.

Companies can use Copilot Studio and Microsoft Foundry to develop agents that work with services such as Teams, Outlook and SharePoint — applications employees already use daily.

 

Agents built through Microsoft Foundry can be published to Microsoft 365 Copilot and Teams, where they become discoverable through their agent catalogues. Microsoft also allows organisations to control whether an agent is available only to its creator or to other employees.

 

Microsoft's advantage lies in distribution. Rather than convincing users to visit a new marketplace, it can place agents inside software already embedded in workplace routines.

 

The wider market

 

Beyond these companies is a less organised layer of community-run directories and development platforms, including MCP hubs, Hugging Face Spaces and Replit's agent marketplace.

These offer developers alternative distribution channels without relying entirely on one large AI company to decide which tools are listed.

Building the checkout counter for AI agents

If AI agents are to complete purchases on behalf of users, the emerging ecosystem also needs payment systems designed for machines acting under human instructions.

 

The payments industry has moved quickly to build this infrastructure.

 

Mastercard introduced Agent Pay in April 2025. The system uses agentic payment tokens intended to identify trusted agents and allow them to make payments under conditions approved by customers.

 

Visa has developed the Trusted Agent Protocol, which helps merchants distinguish legitimate AI agents acting for consumers from malicious automated bots. It has also introduced tools for AI-assisted and machine-based transactions.

 

Stripe's Shared Payment Tokens allow an agent to initiate a payment without receiving the customer's actual card details. Each token can be limited to a particular seller, amount and period.

 

The next step is standardisation.

 

The Linux Foundation announced the x402 Foundation in April 2026 to oversee an open protocol contributed by cryptocurrency platform Coinbase. The foundation became operational on July 14, bringing together companies from the payments, cloud and technology industries.

 

The initiative aims to establish common rules that allow AI agents and applications to exchange value across platforms without depending entirely on proprietary payment systems.

The move follows the same principle that helped MCP gain adoption: competing companies increasingly recognise that common infrastructure can expand the overall market while leaving them free to compete through the products and services built on top of it.

Regulators turn their attention to AI gatekeepers

The comparison with app stores is not merely a convenient description. It is increasingly becoming a question for regulators.

 

Under the European Union's Digital Markets Act (DMA), designated gatekeeper companies face restrictions on favouring their own products, limiting access to users or preventing outside businesses from competing through their platforms.

 

The European Commission has designated companies including Alphabet, Apple, Meta and Microsoft as gatekeepers for specific services. AI assistants have not yet been designated as a separate category of core platform service under the law.

 

That position is beginning to face scrutiny as AI assistants become an important route through which users discover products, services and information.

 

MLex reported in June that European regulators had sought usage data relating to AI assistants developed by companies including Alphabet, Anthropic, Meta, Microsoft and OpenAI as they considered how the services might fit within existing competition rules.

 

A separate case involving WhatsApp has already highlighted the competition concerns surrounding access to large platforms. On June 9, 2026, the European Commission ordered Meta to restore free access to WhatsApp for rival general-purpose AI assistants while it continued an antitrust investigation. The Commission said Meta's restrictions risked causing serious and irreparable damage to competition.

 

The case is being pursued under EU competition law rather than the DMA, but it raises a similar question: whether a company controlling a major distribution channel should be allowed to favour its own AI assistant over rivals.

 

The Commission's first review of the DMA, published on April 28, concluded that the law remained fit for purpose but identified AI and cloud computing as areas requiring closer attention as digital markets evolve.

Where India stands

India does not yet have an ex-ante digital competition law comparable with the DMA. A proposed Digital Competition Bill emerged from recommendations made by a parliamentary panel in 2022. The draft sought to regulate large digital companies before potentially anti-competitive conduct caused market harm, rather than relying only on investigations after the event.

 

The proposal faced resistance from global technology companies and Indian businesses, which argued that its definitions were broad and could bring fast-growing domestic companies within the same framework as established global platforms.

 

The government subsequently indicated that further market studies would be conducted before it proceeded with an ex-ante framework. The Ministry of Corporate Affairs and the Competition Commission of India have continued examining competition in digital markets, but no timeline has been set for introducing a revised Bill.