Skip to content
All posts

Perspectives

Claude's Connector Directory and ChatGPT Apps: The New Discovery Shelf for SaaS

Claude's connector directory and ChatGPT's plugin directory now suggest tools mid-conversation. What SaaS companies should build, list, and measure, and what the platforms have not said.

In this article
  1. Two shelves inside the same assistant
  2. How the Claude connector directory decides what to suggest
  3. What changed in ChatGPT: apps now live inside plugins
  4. Should your SaaS company build one?
  5. What a listing and its tool descriptions need
  6. How to measure the action shelf
  7. What neither platform has said yet
  8. Get on the shelf your buyers already use

The Claude connector directory is Anthropic's catalog of reviewed MCP servers, and every listing in it can be suggested inside a conversation when it fits what the user is trying to do. ChatGPT has its own version: the app directory, now folded into a plugin directory, where ChatGPT can also suggest a plugin mid-conversation.

Picture a customer success lead asking Claude to pull last quarter's churned accounts and draft a win-back email. If your product holds that data and has a listed connector, Claude can offer it right there. If not, your product is never in the room.

This post is for SaaS leaders deciding whether to be on that shelf. MCP, the Model Context Protocol, is the open standard both platforms use to connect an assistant to outside tools and data.

Two shelves inside the same assistant

Buyers meet your product inside an AI assistant in two moments. The first is the answer: someone asks which tool to use, and the assistant names a shortlist drawn from the pages, reviews, and sources it reads. That is the work of generative engine optimization.

The second is the action: someone asks the assistant to do the job, and it reaches for a connector. Our view is that this shelf is becoming a discovery channel next to search, and it deserves its own line in the plan. A connector does not make you the recommended answer, and a strong answer does not get you called when work needs doing.

The shelf is filling up. Anthropic said on April 23, 2026 that since launching in July 2025, the Claude directory has grown to over 200 connectors, available on all Claude plans.

Using connectors in your own marketing work is a separate topic, covered in how to connect Claude MCP to your marketing stack. This post is about being one.

How the Claude connector directory decides what to suggest

As of November 2026, Anthropic's own pages say:

  • Every directory entry is automatically eligible for Suggested Connectors. Custom connectors added by URL are never suggested, and the open MCP Registry does not surface you in Claude; only a directory listing does.
  • Directory ranking is usage-based, "similar to other app stores," per the directory documentation.
  • Claude suggests an app when it matches what the user is asking. When more than one could help, it shows them and the user chooses. Anthropic says there are no paid placements or sponsored rankings.
  • New listings start with a Community label. Anthropic may escalate listings it flags as highly useful to Verified review, and there is no application.

Anthropic does not document how relevance and usage are weighed in a suggestion, or how much the Verified label changes the order.

What changed in ChatGPT: apps now live inside plugins

The "ChatGPT app directory" has moved. OpenAI introduced the app directory on December 18, 2025, and on July 9, 2026 its release notes say it replaced the App Directory with the Plugin Directory, leaving existing app connections unaffected.

In OpenAI's terms, an app connects ChatGPT to a service, and a plugin packages apps, skills, or both. Its plugins help article says the directory is available across ChatGPT plans, that ChatGPT uses installed plugins when relevant, and that people can review a plugin suggested during a conversation.

On ranking, OpenAI says less. Its plugin guidelines say plugins with strong real-world utility and high user satisfaction may be eligible for enhanced distribution, such as directory placement or proactive suggestions.

Should your SaaS company build one?

Build now if three things are true:

  1. Users already ask an assistant to do something with your product's data: pull a report, create a record, update a ticket.
  2. You have an API and can support OAuth 2.0, which Anthropic requires for authenticated services.
  3. The job can finish inside a conversation.

Wait if your value lives entirely in your own interface. Both platforms treat this as a product surface, not an ad slot: Anthropic rejects tool descriptions that promote products, and OpenAI bars plugins from showing subscription plans, promoting upgrades, or advertising pricing.

Both directories take a remote MCP server over HTTPS, so one server can serve both, though the submissions are separate.

What a listing and its tool descriptions need

Both directories ask for similar things.

What each directory asks for, from Anthropic's and OpenAI's submission guides (November 2026)
RequirementClaude connector directoryChatGPT plugin directory
Listing copyName, one-liner, description, categories, iconDisplay name, subtitle, long description, icon, default prompts
Tool rulesA title and a read-only or destructive hint on every toolAccurate tool names and descriptions, plus test cases for reviewers
Ownership proofNone requiredA domain verification challenge
Who submitsAnyone on a paid Claude planA verified individual or organization

The room for copy differs. Anthropic's submission guide allows a name of up to 100 characters, a one-liner of up to 200, and a description of up to 2,000. OpenAI's submission guide caps the display name and subtitle at 30 characters each, allows a 4,000-character description, and asks for five positive and three negative test cases.

Tool descriptions matter more than the listing, because the model reads them when it decides:

  • Say what the tool does and when to use it. OpenAI's metadata guide recommends opening with "Use this when" and naming cases where it should not run.
  • Name tools after the job in your users' words, such as "create invoice."
  • Never tell the model to prefer you. Anthropic rejects descriptions that interfere with Claude calling other tools, and OpenAI bans fields that manipulate plugin selection.

How to measure the action shelf

Anthropic gives listings a dashboard. Its listing documentation shows a directory rank, the Claude accounts that used a tool, tool calls, and an error rate, each over 30 days. The metrics are in beta, can lag by up to 24 hours, and count only traffic from Claude.

OpenAI's metadata guide describes a test worth running on both platforms: a fixed set of direct prompts that name your product, indirect prompts that describe the outcome, and negative prompts that should go elsewhere. Track precision (did the right tool run?) and recall (did it run when it should?).

Then tie it to revenue: new OAuth connections from each assistant, accounts that keep using the connector, and deals where it was the first touch. A connector call never loads your website, so web analytics will miss it.

What neither platform has said yet

In the pages we checked, neither platform reports how often a listing was suggested and not chosen, the closest thing to an impression, and neither publishes a ranking formula. Treat any claimed formula as a guess.

Get on the shelf your buyers already use

If your product does a job people would rather ask for than click through, list it in both directories, write tool descriptions as carefully as your homepage, and measure the action shelf next to your AI answer visibility. The connector earns the call; your content earns the shortlist.

Our SEO and GEO program for SaaS companies works on the answer shelf, where buyers build the list. Talk to our team to see where you show up today.

Frequently asked questions

What is the Claude connector directory?

It is Anthropic's catalog of MCP servers that work across Claude's apps, and every listing can be suggested in a conversation when it fits the user's task. Custom connectors that people add by URL are never suggested.

How does Claude decide which connector to suggest?

Claude suggests a connector when it matches what the user is asking, and directory ranking is usage-based. Anthropic does not document how relevance and usage are weighed against each other.

Is the ChatGPT app directory still available?

OpenAI replaced it with the Plugin Directory in July 2026, and existing app connections were not affected. In OpenAI's terms, an app connects ChatGPT to a service, and a plugin packages apps, skills, or both.

Can you pay to be suggested in Claude or ChatGPT?

Not in Claude, where Anthropic says there are no paid placements or sponsored rankings. OpenAI's plugin guidelines tie directory placement and proactive suggestions to real-world utility and user satisfaction.

Do you need separate builds for Claude and ChatGPT?

Usually not, because one remote MCP server can serve both platforms. The listings and reviews are separate, so plan for two submissions.

How do you measure a Claude connector or ChatGPT plugin?

Use Anthropic's dashboard for directory rank, accounts that used a tool, tool calls, and error rate, and run a fixed set of direct, indirect, and negative prompts to track precision and recall. Then tie new connections and active accounts to pipeline.

Get found and chosen in AI answers.

Book a call and we will show you where Google and AI answers leave you out today, and what we would ship first to get you in.