Search "Claude plugin" and you'll find four different products under one name. A Claude plugin is an installable bundle of skills, connectors, slash commands and agents for Claude Code or Cowork. A Claude connector is a single MCP integration that lets Claude read from and act in another service, such as Shopify, Google Drive or Slack. ChatGPT apps do the connector's job on OpenAI's side and also run on MCP, while Meta Muse and OpenAI Dots are personal agents that go out to the web and act for you.
You add a plugin from one directory, or with the /plugin command in Claude Code. One of the four products isn't Claude's at all. It's ChatGPT plugins, which OpenAI retired in 2024.
Meta Muse and OpenAI Dots try to buy things for people. They are neither plugins nor connectors.
This guide is current as of November 2026. It covers which one you need and how to install it. It also covers what each route means for agentic commerce if you sell online.
Connectors, plugins and apps bring tools into an assistant; Meta Muse and OpenAI Dots go out to stores and checkout.
"Claude plugin" can mean four different things
Find your row and jump to it.
Routing table, as of November 2026
If you mean… | It's called… | Where it lives | Jump to |
|---|---|---|---|
A bundle for the terminal coding tool | Claude Code plugin | Claude Code, via | What is a Claude plugin? |
Something you add in the Claude apps or Cowork | Connector or plugin | The Customize directory | Connectors and plugins |
Claude inside VS Code or JetBrains | Editor extension | The IDE marketplaces | Anthropic's VS Code extension and JetBrains plugin |
ChatGPT plugins | Retired in 2024 and replaced by apps. Since July 2026, the name covers bundles of apps and skills (gradually.ai) | ChatGPT | What happened to ChatGPT plugins? |
Anthropic lists skills, connectors and plugins in one directory. That's why non-developers can't tell them apart.
The editor extension is a separate product. It puts Claude in your IDE. In VS Code, typing /plugins in the prompt box opens Manage plugins (Claude Code docs).
To find a specific vendor such as Adobe, search the directory for the company name. Adobe for Claude, for example, is a plugin. It connects Claude to Acrobat, Adobe Express, Photoshop and other Adobe apps (Adobe Help Center).
What is a Claude connector?
A connector lets Claude read data from another service. Where you allow it, Claude can also take actions there. Connect Google Drive and Claude can quote a document. Connect Shopify and it can look up orders and products.
You install Shopify's connector from the Claude directory. Then you approve its data access in your Shopify admin (Shopify Help Center).
Connectors run on the Model Context Protocol (MCP). MCP is an open standard for how an AI assistant finds a service's tools and data and calls them. A service builds one MCP server, and any assistant that speaks MCP can use it.
Read actions, write actions and custom connectors
"Find last week's orders over $200" is a read. "Update the description on this product" is a write. In a live store, a write changes what customers see.
Tool permissions exist for that reason. You can set each tool to Always allow, Needs approval or Blocked (Anthropic connector docs).
Directory connectors come from the listing. A custom connector is the URL of a remote MCP server that you or a vendor runs. Every plan from Free to Enterprise can add one (Claude Help Center).
How to add a connector
In the Claude app, open Customize, select Connectors, then Discover.
Search for the service, for example Google Drive.
Select Connect to Claude, sign in, review the access requested and approve it.
Under Tool permissions, set write tools you don't want to Needs approval or Blocked.
In a chat, turn the connector on from the + menu. Ask a question that needs the data, and check the answer against the source.
For a custom connector, select Add custom connector and paste your remote MCP server URL.
On Team and Enterprise plans, an Owner adds the connector first. Members then connect their own accounts (Anthropic connector docs).
If you run a store, the protocol also works in the other direction. Opttab's AI Commerce includes an MCP server. AI agents can use it to search your products, check stock and apply your policies.
In Customize > Connectors, use Add custom connector to paste a remote MCP server URL.
What is a Claude plugin?
A plugin is a package. What it can do depends on what's inside.
One install can bundle skills, connectors (MCP servers), slash commands, subagents and hooks. Some plugins also add LSP servers. These give Claude code intelligence for a programming language.
Plugins work in Claude Code and Cowork. Their skills, commands and connectors also work in chat on claude.ai and the desktop app (Anthropic plugin docs). Engineers aren't the only audience. Anthropic's knowledge-work plugins include marketing and sales plugins.
Anthropic opened a plugin submission portal on 25 September 2026 (Unite.AI). Developers on paid Claude plans can submit. Anthropic checks each version before it goes live (Anthropic).
Inside a Claude plugin: one manifest, five component types, and which of them can run code.
Plugin components at a glance
Check the components first. They tell you whether a plugin only adds instructions or also runs code and reaches other systems.
Plugin components, as of November 2026
Component | What it does | When it runs | Runs code or reaches external systems? | Example |
|---|---|---|---|---|
Skill | A SKILL.md file of instructions, plus optional resources | When a task matches its description | Only if it bundles scripts | Brand voice rules for product copy |
Connector (MCP server) | Links Claude to a service's tools and data | When Claude calls one of its tools | Yes. Local servers also run on your machine | A Shopify connection |
Slash command | A saved prompt you trigger with | When you type it | Not by itself | Draft release notes |
Subagent | A specialised agent with its own instructions and tools | When it's handed a task | Through any tool it's granted | A code reviewer |
Hook | A script that fires on lifecycle and tool-use events, such as | On those events (hooks reference) | Yes. Command hooks run shell commands, and other types can call a URL, an MCP tool or a model | Run a linter after Claude edits a file |
LSP server | Adds go-to-definition, diagnostics and type information | In the background on code | Yes, locally | A TypeScript language server |
How to install a Claude plugin
In Claude Code:
Open the plugin menu and browse the official marketplace in the Discover tab.
/pluginOr install a plugin by name.
/plugin install plugin-name@marketplace-nameAdd a third-party marketplace hosted on GitHub.
/plugin marketplace add owner/repoRefresh a marketplace for new plugins and versions.
/plugin marketplace update marketplace-namePause or remove a plugin.
/plugin disable plugin-name@marketplace-name
/plugin uninstall plugin-name@marketplace-name
Inside a session, /plugin install opens the plugin's details first. You can review it there and pick user, project or local scope (Claude Code docs).
The /plugin install flow in Claude Code: browse Discover, review the plugin, then choose a scope (illustration, not a screenshot).
In the Claude apps and Cowork:
Open Customize and select Plugins.
In Discover, choose a plugin and read what it includes.
Select Add. Disable or remove it later from its menu.
A plugin added there is saved to your account. It's then also available in Cowork and Claude Code (Anthropic plugin docs).
The Claude plugin marketplace
A marketplace is a catalogue Claude Code can install from. Anthropic runs the official one. Anyone can publish another by adding a marketplace.json file to a GitHub repository.
That makes the ecosystem large and unvetted. ClaudePluginHub, a third-party index that pulls plugins from across GitHub, listed 115,005 plugins on 10 November 2026 (ClaudePluginHub). It ranks them by what developers copy. A high rank reflects popularity, and says nothing about safety.
Start with the official Claude directory. Anthropic checks and safety-scans every submission, and lists a plugin only after it's approved (Anthropic). Open-source marketplaces on GitHub come next, but there you do the vetting. To roll plugins out to a team, run a private marketplace with an approved list.
Before you install a third-party plugin
Read the manifest (
plugin.json) to see every component it adds.Open any hook scripts, because hooks run shell commands.
Check what each MCP server connects to and whether it can write.
Check the marketplace source, the maintainer and whether the code is public.
Pin a version so an update can't change behaviour unnoticed.
Remove plugins you've stopped using.
Every MCP tool a plugin adds has a definition that takes context space. In Claude Code, tool search is on by default. Only tool names and server instructions load when a session starts. Full definitions load when Claude needs them (Claude Code docs). Tools set to load upfront still use space, so run /context to see what's loaded before you add more.
Team rollout, scopes and update policy each get their own post. So does building your own plugin with an annotated plugin.json.
Claude connector vs Claude plugin
Connector | Plugin | |
|---|---|---|
What it is | One MCP integration with one service | A package of skills, connectors, commands, subagents, hooks and LSP servers |
Who it's for | Anyone who wants Claude to see or act in a tool | Anyone who wants a repeatable workflow |
Where it runs | Claude apps, Cowork and Claude Code | Claude Code and Cowork. Skills, commands and connectors also work in chat |
What it can do | Read data and, where allowed, act in that service | Whatever its components do, including running scripts |
How you get it | The Customize directory, or a custom MCP server URL | The Customize directory, or |
Runs code on your machine? | Only local MCP servers | It can, through hooks, scripts, local MCP servers and LSP servers |
Add a connector when you need Claude to see your store or your files. Install or build a plugin when you need a repeatable workflow that uses that connector.
What happened to ChatGPT plugins?
OpenAI's route has changed shape several times. Here's the sequence as of November 2026.
2023: ChatGPT plugins launch, letting ChatGPT call third-party services.
2024: OpenAI retires plugins. Plugin conversations stopped on 9 April 2024 (OpenAI Developer Community).
GPTs: custom versions of ChatGPT with their own instructions and files, rather than integrations.
Apps in ChatGPT: OpenAI introduced apps on 6 October 2025. They run inside a conversation. Developers build them on MCP with the Apps SDK. Today they're submitted and listed as plugins, in one directory that ChatGPT and Codex share (OpenAI).
September 2025: OpenAI and Stripe launch the Agentic Commerce Protocol. Instant Checkout lets shoppers buy inside ChatGPT.
March 2026: Instant Checkout is retired, and merchants run their own checkout (OpenAI).
July 2026: the plugins name returns for bundles of apps and skills (gradually.ai).
For builders, one change matters most. ChatGPT apps (now listed as plugins) and Claude connectors use the same protocol. Both build on MCP and support MCP Apps, the shared extension for interactive UI in chat (OpenAI, Anthropic).
OpenAI tells developers to start with the MCP Apps spec. ChatGPT-only features go in an optional window.openai layer (OpenAI). Build to the shared spec and one MCP server can back both assistants.
Packaging and review still differ. ChatGPT takes a plugin ZIP through the OpenAI Platform (OpenAI). Claude takes a remote MCP server URL or a GitHub bundle through its own submission portal (Anthropic). Test the server in both before you submit.
How OpenAI’s and Anthropic’s integration and commerce routes evolved, as of November 2026.
Claude vs ChatGPT: how the two routes compare
The two routes share plumbing. They differ on discovery and checkout.
Claude and ChatGPT integration routes, as of November 2026
Claude (connectors and plugins) | ChatGPT (apps) | |
|---|---|---|
Integration standard | MCP (Anthropic) | MCP, through the Apps SDK (OpenAI) |
Where users find them | The Customize directory, and | The Plugins tab in ChatGPT on the web and desktop, one directory shared with Codex (OpenAI) |
Who builds them | Anthropic, vendors, and anyone with a marketplace. Official listing goes through the submission portal | Developers with a verified OpenAI Platform organization. Plugins pass automated checks, then a review team tests them with set test cases before the developer can publish (OpenAI) |
Checkout | Claude Commerce Agents keep checkout on the merchant's site (Anthropic) | Instant Checkout retired in March 2026, so merchants run checkout |
How users confirm actions | Per-tool settings: Always allow, Needs approval or Blocked | The developer labels each tool, for example as read-only or able to delete data. ChatGPT uses the labels to frame each call and asks for explicit approval before tools that may delete or overwrite data (OpenAI) |
Best for | Working across your own files and tools, and team workflows | Reaching people while they chat in ChatGPT |
For brands, one well-built MCP server and one clean product feed should cover both routes. Confirm that against the Apps SDK docs before you plan engineering work around it.
Anthropic's commerce answer: Claude Commerce Agents
Anthropic released Claude Commerce Agents on 2 September 2026. It's an open-source blueprint for shopping and merchant agents, under the Apache 2.0 licence (Anthropic). The agent runs on the merchant's own site, and checkout stays with the store. Shopify support arrived the same day.
If you sell on Shopify, read the blueprint before you commission a custom shopping assistant.
The new wave: personal agents that shop for you
Connectors and apps wait for you to call them. Personal agents take a goal and act on it. They often act on websites that never agreed to an integration.
Meta Muse
Meta launched Muse on 8 September 2026. It browses retailers and buys after the user clicks a final "Confirm purchase" button, according to CNBC. Its partners include Stripe, Shopify, Walmart, Best Buy and Sephora.
Amazon blocked Muse over how the agent identifies itself and handles credentials, GeekWire reported. Meta disputes Amazon's account.
OpenAI Dots
OpenAI launched Dots on 29 September 2026 at DevDay. Each Dots agent is always on and has its own cloud computer and browser, according to CNBC.
Dots is available to Pro and Business Premium users in eligible markets, according to TechCrunch.
A dot can research on its own, before you ask. In that mode, its tools can only read permitted connected sources and save private notes. They can't send messages, change content through plugins or control a browser. That's according to OpenAI's help center, as reported by Search Engine Journal.
What this means for your site
The Amazon and Muse standoff shows that agents now arrive at your site as visitors. Not all of them say who they are. Before you set a policy for agent traffic, find out what's already arriving.
Opttab's agent and website analytics shows which AI bots and crawlers visit your site and what they read. It also shows the human traffic and conversions from AI search, through a site tag or GA4.
Tools you bring in vs agents that go out
Most questions about control and risk come down to which lane you're in. Connectors, plugins and apps are tools you bring into an assistant. Muse and Dots go out and act for you.
Tools vs agents, as of November 2026
Connectors, plugins and ChatGPT apps | Meta Muse and OpenAI Dots | |
|---|---|---|
Who acts | You, through the assistant | The agent, on its own between instructions |
Where it acts | Inside the assistant, calling the service's API through MCP | On retailer websites, in a browser |
Permissions | Granted per connector or app, usually by signing in | Granted per agent, and limited by sites that block agents |
Who holds payment details | In ChatGPT, mostly the merchant. Plugins send buyers to the merchant's own checkout or show cards already saved with that merchant; ChatGPT's own payment sheet is a beta for select marketplaces (OpenAI). Anthropic's Claude Duet pays through Stripe's Link with a virtual card, after you approve each purchase (Stripe) | Muse uses Stripe's Link: the card saved in Link, or a single-use virtual card scoped to the approved purchase (Stripe). A dot uses cards you've saved on a merchant's website, and each purchase needs your approval, which you can give in advance for a specific purchase (OpenAI) |
What can go wrong | An unintended write action, or plugin code you didn't read | The wrong item, price or currency, or a site misreading the agent |
What the merchant sees | An API call to its MCP server | A browser visit that may or may not identify itself |
What this means for shoppers
Before you let any assistant or agent buy for you:
Know which confirmations it asks for, and at which step.
Check which accounts you connected, and whether each has read or write access.
Check where your card is stored and who processes the payment.
Read the returns policy on the store's own site.
Check the final price and currency before you confirm.
Revoke access you no longer use.
Steps 4 and 5 matter most, because AI answers often leave those details out. Opttab Research ran an agentic commerce study with data collected on 9 November 2026. Only 23.8% of answers gave a delivery or returns detail. 54.4% stated no concrete price.
That study covered logged-out ChatGPT, Gemini, Google AI Mode and Bing AI Mode. It did not cover Muse or Dots.
What this means for brands in agentic commerce
For a logged-out shopper, agentic commerce today is a referral business. In our agentic commerce study, we asked ChatGPT, Gemini, Google AI Mode and Bing AI Mode 60 buying questions in six countries. Each question was in the local language. We read all 1,440 answers on 9 November 2026.
0 of 1,440 answers completed a purchase.
2.9% offered checkout in chat.
71.7% of "buy it for me" answers said the assistant couldn't buy.
69.6% of answers pointed the shopper to a store in text.
That 71.7% sits between our Black Friday 2026 (64.4%) and Cyber Monday 2026 (82.1%) results. Prompt sets, surfaces and dates differ, so read the comparison as directional. The study didn't test Claude, Muse or Dots. It describes the default logged-out experience, not signed-in or connected agents.
The moment that counts is the hand-off. When an assistant sends a shopper to a store, you want it to send them to yours. Three jobs get you there.
Only 9.4% of AI buying answers named a brand’s own store first (Opttab Research, agentic commerce study, 9 November 2026).
Job 1: Be reachable by tools (an MCP server for ecommerce)
Run an MCP server so Claude connectors, ChatGPT apps and agents can query your live products, stock and policies. Without one, an assistant only knows what it last read about you.
Keep your product feed clean for ACP and UCP too. Shopping surfaces read feeds when they don't call you directly.
Most stores need neither a plugin nor an app to start. Both routes sit on MCP, so a working MCP server and a clean feed come first.
In Opttab: AI Commerce gives you an MCP server and the Feed Optimizer. The Shopify integration connects your catalog.
Job 2: Be chosen by agents
In the same study, only 9.4% of answers named a brand's own store first. Multi-brand retailers came first in 37.8% of answers. Marketplaces came first in 23.0%.
Assistants quote what they can read. Put price, delivery, returns and stock facts in plain text. Name your market and currency.
Market matters more than most brands assume. Only 27.1% of answers to Canadian shoppers named a Canadian store first. In the US, it was 72.9%.
Surfaces differ too. ChatGPT showed product cards in 44.7% of answers. Google AI Mode showed them in 27.5%. Gemini and Bing AI Mode almost never did.
Answers also shift from run to run. In Opttab Research's Black Friday 2026 study (data collected 31 October 2026), two answers to the same question shared a top pick only 28.4% of the time. Track your share across many prompts, surfaces and markets rather than one position.
In Opttab:
Generate your customers' buying prompts with the Prompt Generator.
See how often those prompts are asked, and from which countries, in Prompt Volume.
Track how often you're named across AI models, including Claude, by country and language in AI visibility tracking.
Find gaps against the most-cited pages with GEO/AEO actions.
Write missing policy and FAQ content in Content Studio.
Serve clean, structured pages to AI bots with AXP.
Job 3: Own the checkout
Instant Checkout is retired. Claude Commerce Agents keep checkout on the merchant's site. Either way, the hand-off lands on your checkout. Make that path work for a referred shopper and for an agent visiting on someone's behalf.
In Opttab: run the Agentic Commerce Readiness Audit in AI Commerce for a 0–100 score and a roadmap. Then use agent analytics to see which agents reach your site and which visits convert.
How to get started this week: a checklist for online stores
Run a free AI Visibility Report to see what AI says about your brand today.
Add the Shopify connector in Claude. Ask it three questions about your catalog, and note what it gets wrong.
Run the Agentic Commerce Readiness Audit and work through its roadmap.
Fix feed quality in the Feed Optimizer.
Stand up an MCP server through AI Commerce.
Put returns windows, delivery cut-offs and prices in plain text on one crawlable page. Content Studio can draft it from your visibility gaps.
Set up tracked buying prompts per market in AI visibility tracking.
Check agent analytics weekly for new AI bots and the pages they read.
How to do this in Opttab
Readiness: the Agentic Commerce Readiness Audit scores your store from 0 to 100 with a roadmap.
Reachability: AI Commerce includes an MCP server and the Feed Optimizer.
Tracking: AI visibility tracking follows your buying prompts by model, country and language.
AI Commerce is part of the Growth plan. See pricing, which includes a 7-day free trial. To weigh Opttab against other tool categories, see how Opttab compares.
FAQ
How do I install a plugin in Claude?
In Claude Code, type /plugin and browse the official marketplace, or run /plugin install plugin-name@marketplace-name. In the Claude apps and Cowork, open the Customize directory, choose a plugin and select Add. For third-party plugins, read the manifest and hook scripts first.
Is a Claude connector the same as an MCP server?
They're closely linked. The MCP server is the software on the service's side, and the connector is how Claude links to it and signs in. A plugin can bundle one or more MCP servers with skills and commands.
Are ChatGPT plugins still available?
The original ones aren't. OpenAI retired them in 2024, and apps in ChatGPT, built on the Model Context Protocol, took over that role. Since July 2026, ChatGPT has used the plugins name again for bundles of apps and skills.
Can Claude buy things for me?
Claude can act in a store you connect if the connector allows write actions. Anthropic's Claude Commerce Agents blueprint runs on a merchant's own site, and the store keeps checkout. Opttab Research didn't test Claude, but across four other assistants, 0 of 1,440 logged-out answers completed a purchase.
What is OpenAI Dots?
Dots are always-on personal agents that OpenAI launched on 29 September 2026 at DevDay. Each has its own cloud computer and browser, according to CNBC. TechCrunch reports they're on the Pro and Business Premium plans, and CNBC reports that connected apps are read-only by default.
Is Meta Muse safe to shop with?
Muse asks for a final "Confirm purchase" click before it buys, according to CNBC. Amazon blocked Muse over agent self-identification and credential concerns, and Meta disputes that account. Before you confirm, check the final price, currency and returns policy on the store's own site.
See what AI tells your shoppers
Find out what Claude, ChatGPT and other AI assistants say about your store. Get a free AI Visibility Report, no credit card needed.
Then act on it. Run the Agentic Commerce Readiness Audit in AI Commerce, start a 7-day free trial, or book a demo.
