TL;DR: The large assistants turned into distribution platforms between late 2025 and 2026. OpenAI launched apps in ChatGPT in October 2025 and moved them into a plugin directory shared with Codex in July 2026; Anthropic's Claude Marketplace lists more than 2,000 plugins and connectors (company-reported); Google opened a preview of Connected Apps for Gemini and Search; Microsoft 365 Copilot replaced its Agents page with a plugin registry. Most of these run on the Model Context Protocol, and MCP Apps lets one interface render inside several assistants. The PM decision is whether the jobs your users bring to an assistant are ones your product should meet there, under platform rules you do not control, and how you keep the customer relationship when the conversation happens somewhere else.
How the platforms work, November 2026
Each row is from the platform's own announcement or documentation. Catalogue sizes are company-reported.
| Platform | Listing | Built on | Notable rules and facts |
|---|---|---|---|
| ChatGPT and Codex | Plugin directory (the app directory until 9 July 2026) | Apps SDK on MCP; fully compatible with MCP Apps since February 2026 | Physical goods only for commerce; no ads; no ChatGPT-specific fees; Instant Checkout in beta for select partners |
| Claude | Claude Marketplace (September 2026), connectors directory since July 2025 | MCP connectors, MCP Apps interactive connectors, Agent Skills | More than 2,000 plugins and connectors; buyers can apply part of their committed Anthropic spend |
| Gemini and Search | Connected Apps (preview, October 2026) | Cloud-hosted MCP server or Android AppFunctions | Developers choose Search, the Gemini app or both; more than 50 Connected Apps |
| Microsoft 365 Copilot | Plugin registry (September 2026), Agent Store since May 2025 | Skills, connectors and MCP configuration in one plugin | More than 100 plugins |
Two patterns stand out. The unit of distribution converged on a plugin: a bundle of an MCP server, skills and optional interface. And because MCP Apps became an official MCP extension in January 2026, supported by ChatGPT, Claude, Goose and VS Code, one well-built server and interface can reach several assistants with modest per-platform work.
From server to user
Follow the flow, one stage at a time.
Illustrated sequence; animation speed does not represent processing time.Read diagram description
A company exposes its capabilities as an MCP server, adds an interactive interface with MCP Apps, and packages them with skills as a plugin. It submits the plugin for each assistant's review, gets listed in the directory, and may be suggested inside conversations. The user acts in the chat, and transactions hand off to the company's own site or app. Utility and satisfaction earn more placement.
The rules shape the business model
The platforms decide what you may sell, how you may be found and what you may say. OpenAI's plugin guidelines, as of November 2026, allow commerce "only for physical goods" and bar selling digital products, subscriptions or credits; prohibit advertisements and ChatGPT-specific fees; and forbid instructing the model to prefer one plugin over others. Placement is earned: plugins with "strong real-world utility and high user satisfaction may be eligible for enhanced distribution opportunities, such as directory placement or proactive suggestions".
So a subscription product cannot simply sell its plan inside ChatGPT today. OpenAI's own early guidance points the same way: developers link out to their sites or apps to complete transactions for physical goods. Zillow's launch app shows listings in the chat, then sends users back into Zillow to tour a home, contact an agent or explore financing.
Disintermediation is the trade
Inside an assistant, the user may never see your brand's home page, your recommendations or your upsells. You gain reach and lose context. Weigh it explicitly:
- What you gain. Presence where users already ask; OpenAI's launch post cited reach to over 800 million ChatGPT users (company-reported). Stack Overflow, in an OpenAI customer story from October 2026, said its registrations "tripled overnight" after launching a ChatGPT plugin (company-reported, no base given).
- What you risk. The assistant owns the conversation, the user's history and the ranking among competitors. The platform can change the rules: OpenAI turned its app directory into a plugin directory about seven months after opening it.
Expose the actions that bring users to you, such as search, quotes and availability, and keep the ones that build the relationship, such as accounts, loyalty and payment, on your own surfaces where the rules allow.
What to measure
Treat each assistant as a channel with its own funnel: plugin installs or connections, sessions that use your tools, actions completed in chat, hand-offs to your site or app, and conversions after the hand-off. Compare cost per acquired customer and retention against your other channels, and watch whether assistant users later come back directly. Track your placement too: directory position and how often you are suggested, since OpenAI's guidelines say utility and satisfaction may earn placement.
In the interview
Practise prompts such as "should we build a ChatGPT app?", "a competitor launched in Claude; do we follow?" or, at a platform company, "design the rules for third-party apps in our assistant". A weak answer builds on every platform because it is new. A Hire answer picks a platform by audience. A Strong Hire answer starts from the jobs users bring to assistants, builds once on MCP and MCP Apps, reads each platform's commerce and discovery rules, decides which actions to expose and which to keep, measures the channel like any other, and plans for the rules changing.
Practise with building inside an AI assistant.
Next: product strategy and distribution and platform risk and lock-in.