How to get your MCP server into every AI app store (Nov 2026)

I recently built Wherbi, a shopping MCP server. Its one job: you ask an AI where to buy something, and Wherbi answers with shops and prices. “Where to buy it? Just ask.” When someone buys through one of its links, Wherbi earns an affiliate commission.

Building the server was the easy part. Getting it in front of people was not. Every AI app now has its own store, its own form and its own rules. This post is the map I made while working through them. All facts are as of 4 November 2026. These stores change every month, so check the date before you rely on a detail.

Why this matters now

A year ago you installed an MCP server by pasting JSON into a config file. Only developers did that.

That’s changing. The labs now show MCP tools inside the chat itself. ChatGPT, Claude, Gemini and Copilot each have a picker where people browse tools, and the model can also suggest a tool on its own when it fits the question. So every store is a discovery channel. Not being in one means that whole group of users never sees you.

The store list in this post started with a thread by Elliot Garreffa on X (2 November 2026), who mapped the stores first. I checked each one for Wherbi and added what I ran into. Wherbi is live in the official registry; the other stores are in progress.

The stores at a glance

Now the details that matter for each one.

1. The official MCP Registry: start here

The official registry lives at registry.modelcontextprotocol.io. Publish here first, for two reasons. It’s the only one where you control the timing. And several directories, among them Glama, PulseMCP and mcp.directory, copy their listings from it. One publish, several listings.

You describe your server in a server.json and publish it with the mcp-publisher CLI. The registry wants proof that you own the namespace. For a domain namespace you can do that over HTTP: put a public key at /.well-known/mcp-registry-auth on your domain and sign with the private key. No DNS change needed, which is handy when someone else manages your DNS.

Wherbi has been live there as com.wherbi/shopping since 3 November 2026. The reverse-domain name is the namespace you proved.

2. ChatGPT apps

You submit at platform.openai.com/apps-manage. This is the strictest review of the big ones:

• Organisation verification with an ID. Do it first, it can take a while.
• A demo video of the app working inside ChatGPT.
• Test cases: 5 prompts where your app should do its job, and 3 where it should stay out of the way. The negative ones are easy to forget, and they’re the point: OpenAI wants to see that your tool doesn’t grab every question.
• Listing rules. The listing text may not advertise prices, discounts or comparisons. For a shopping tool that’s awkward: describe what it does (“finds shops that sell a product”), not what people save. And the name may not end in “MCP”.

3. Claude’s connector directory

You submit at claude.ai/directory/manage. Any paid Claude plan can submit. A new connector starts as Community: it passes automated checks and appears in the directory.

Verified is a separate tier. Anthropic grants it; there’s no form to apply. When nobody asks for you by name, Claude tends to pick only verified connectors. A Community listing gets you found by people who search for you, much less by people who don’t know you exist yet.

4. Google Connected Apps (Gemini and AI Mode)

Google opened Connected Apps on 30 October 2026. It works in Gemini and in AI Mode in Search. There are only a few apps so far and new ones go through a waitlist. Join it now. A store with few apps is the easiest place to stand out.

5. Meta Muse and Grok

Neither has an open door. Meta adds connectors to Muse from its own catalogue, and you can’t submit. Grok lists about 644 connectors, but I found no public form. For both, the best move today is to be in the official registry and the big directories, so they can find you when they look.

6. Microsoft 365 Copilot Agent Store

Copilot doesn’t list bare MCP servers. You wrap the MCP server as an agent and submit that agent in Partner Center. It needs authentication. That makes sense for work tools, less for a public consumer tool without accounts.

7. The B2B and developer stores

Three more stores are worth knowing, but they suit B2B and developer tools better than a consumer MCP like Wherbi:

• Salesforce AgentExchange. AppExchange and the Slack Marketplace merged into it in April 2026. The place for CRM and Slack tools.
• AWS Marketplace, “AI Agents & Tools”. Free listings are allowed, so you don’t need a paid product.
• Cursor Marketplace. One-click install for developers, and every plugin is reviewed by hand.

Listed is not discovered

After a week of forms it’s tempting to think you’re done. You’re not. Being listed means people can find you if they search for your name. Being discovered means the model picks your tool when someone asks a question you can answer, without knowing you exist.

That second part depends on trust, and the stores show trust with tiers. Claude has Community and Verified. ChatGPT’s strict review works as a gate up front. Copilot needs auth. The pattern is the same everywhere: the model leans on tools the platform vouches for.

So treat a basic listing as step one. What moves the needle is the higher tier, and you usually can’t ask for it. You earn it with a tool that answers well, fast and without errors. Which brings me to the work that never stops.

Agent Tool Optimization

SEO is about ranking in a search engine. This is the same idea for agents: making the model choose your tool, and making the answer good when it does. Some people call it Agent Tool Optimization. It comes down to three steps and a weekly habit.

1. Write down the prompts you want to win. Ten to twenty is enough. Real questions, in the words people use: “where can I buy a Stokke Tripp Trapp in the Netherlands?”, not “product availability lookup”. Add a few that should not trigger your tool. Then run all of them in every agent you’re listed in and note what happens. Was your tool called? Was the answer right? That’s your baseline.

2. Rewrite your tools for those prompts. The model only sees your tool’s name, description and input schema. That’s your whole pitch.

• Give tools neutral, plain names that say what they do. find_shops_for_product beats a clever brand name.
• Write the description for the model: when to use the tool, when not to, and what comes back.
• Keep the schema small and clear. Name parameters the way people think, and give examples.
• Mark read-only tools with the readOnlyHint annotation. A model is more willing to call a tool it knows won’t change anything.

3. Cut latency and errors. A tool that’s slow or fails gets dropped, by the agent in the moment and by the platform over time.

Here’s what that looked like for Wherbi. One tool searched about 30 shopping sites at once and merged the results. On a warm run that was fine. But the first call after a quiet period missed the agent’s time budget of about 3.5 seconds. The agent didn’t wait. It just saw an empty answer and moved on, and from the outside it looked like Wherbi knew nothing. The fix was to stop doing the work at request time. I precompute the data, so the tool now reads a ready answer. It responds in about 0.2 seconds, first call included.

4. Re-run the prompts every week. Models change, stores change, and your tool changes. The same list of prompts, every week, in every agent. It’s the only way to know whether a change helped.

Checklist

Copy this and work through it top to bottom. Or tick it off in the MCP store checklist , which remembers what you’ve done.

Before you submit

• Server runs over HTTPS at a stable URL
• Tool names are neutral and say what they do
• Every read-only tool has readOnlyHint
• Descriptions say when to use the tool and when not to
• First call answers well within 3 seconds, cold start included
• A list of 10-20 target prompts, plus a few that should not trigger the tool

Official MCP Registry

• server.json written
• Public key at /.well-known/mcp-registry-auth
• Published with mcp-publisher
• Listing shows up in Glama, PulseMCP and mcp.directory afterwards

ChatGPT apps

• Organisation verified with ID
• Demo video recorded
• 5 positive and 3 negative test cases written
• Listing text without prices, discounts or comparisons
• Name doesn’t end in “MCP”

Claude connector directory

• Submitted from a paid plan
• Listed as Community

The rest

• On the Google Connected Apps waitlist
• Copilot: only if your tool has auth and suits work
• AgentExchange, AWS Marketplace or Cursor: only for B2B or developer tools

Every week

• Run every target prompt in every agent
• Note which prompts call your tool and which answers are right
• Fix one name, description or slow path, then measure again

What I’d do differently

I’d start with the prompts, not the forms. The stores are a one-time job you can finish in a week. Whether an agent picks your tool is the part that decides if anyone uses it, and you can test that from day one.