Guide · ChatGPT apps

How do users and AI find apps inside ChatGPT (Apps SDK, the plugin directory)?

Updated 2026-11-12 · kagentx

Short answer

Users find apps in ChatGPT by two routes: the official directory (the docs now call it the Plugins Directory, shared by ChatGPT and Codex), and the model deciding in a conversation whether to call your app based on the name, description and parameter docs you wrote. To get listed you upload a ZIP, verify your domain, connect and scan the MCP server, submit for review, and publish after approval. OpenAI has not published directory ranking rules. Test on 2026-11-12: chatgpt.com/apps is an empty shell to crawlers and only a few partner products have introduction pages in the HTML, so your own site needs a page that explains the app too.

What are ChatGPT's apps called now, and are they the same thing as the GPT Store?

No. The GPT Store holds GPTs built from instructions and configuration; this guide covers apps built from an MCP server, skills, or both, which users connect in ChatGPT and whose tools the model then calls. One naming point: OpenAI's docs still live at developers.openai.com/apps-sdk, but the page titles and body text we read on 2026-11-12 say Plugin and Plugins Directory, a directory shared by ChatGPT and Codex, and the changelog mentions both "Apps SDK integrations" and "ChatGPT apps". The docs never say in one sentence whether this is a rename or two separate things; we follow the docs' wording and treat them as one system, which is an inference.

A GPT in the GPT StoreThe apps in this guide (called plugins in the docs)
What it is made ofInstructions, knowledge files and configurationAn MCP server (live data, authentication, tools), skills, and optional UI
How users get itPick it in the GPT StoreInstall and connect it in the Plugins Directory; the model decides when to call it in a conversation
Review required?Not checked in this guideYes: upload a ZIP, automated scans, human review, then you publish after approval
For web crawlersWe tested /gpts and /g/ pages on 2026-11-09: they return the app shellSee the test below

For the GPT Store test and directory submission advice, see Can the GPT Store and AI agent directories get you recommended by AI?; for how an MCP server itself gets into connector directories, see How do you get an MCP server recommended by AI and listed in connector directories?.

How do users find an app?

The official docs describe two routes. One is the directory: "the plugin directory helps people discover and use it in ChatGPT and Codex". The other is the model deciding inside a conversation: the metadata guide says "ChatGPT and Codex decide when to call your tool based on the metadata you provide", so names, descriptions and parameter docs should be written like product copy. The guidelines also say strong plugins "may be eligible for enhanced distribution opportunities, such as directory placement or proactive suggestions", but they give no ranking rule and no threshold for what counts as strong; unverified.

We found no official description of how directory search is ranked. What we can confirm is which fields you fill in at submission (next section); for keywords the docs say only "optional discovery terms".

What fields does a directory listing have, and what are the rules?

FieldOfficial statement (read 2026-11-12)How to write it
NameAvoid generic single words; do not append MCP, MCP Server or Plugin to a company or product nameUse the brand name as it is
description, longDescriptionClear and accurate; no comparing with or disparaging other products, no unverifiable claims, no advertising of price, subscriptions, discounts or promotionsWrite what it can do and which data it reads; no winners and losers, no prices
keywordsOptional array of discovery termsUse the phrases users would use, for example notes, search
categoryRequired; use a category title from the dashboard, such as Productivity or Developer ToolsPick the closest one
capabilitiesList capabilities, for example Search notes, Read notesOne action per entry
Example promptsRequired, to help users get startedWrite 3-5 realistic phrasings, both direct and indirect
Privacy PolicyA clear, published privacy policy is requiredData categories, purposes, retention periods and the controls users have
TranslationsMay be supplied in the package; importing them does not yet change the text shown in the Plugins DirectoryThe language shown in the directory is up to OpenAI; whether this has changed since is unverified

One more fairness rule: no model-readable field (title, description, tool annotation) may steer the model to prefer you over another plugin. Treat it as: only write what you can do yourself.

What steps does review involve?

  1. 1. Verify identity and upload the ZIP

    Pick an organization and project, complete individual or business verification, upload the ZIP, and the system checks metadata and skills. Organization owners can submit; other members need Apps Management Write permission.

  2. 2. Connect and scan the MCP server

    The server needs a stable public HTTPS address. For domain verification you serve a plain-text token at https://<your domain>/.well-known/openai-apps-challenge. One plugin can connect only one MCP server.

  3. 3. Prepare the review materials

    If sign-in is needed, provide a test account with sample data that needs no new sign-up or second-factor code. The docs ask for 5 positive test cases, 3 negative ones (prompts that should not trigger), a walkthrough video and release notes.

  4. 4. Submit, wait for the email, then publish yourself after approval

    Only one review can be active per plugin. After publication, changes to the MCP server are scanned automatically; changes to metadata or skills need a new ZIP.

How do I make the model more likely to pick my tool?

The official metadata guide says: first write a labelled prompt set (direct prompts that name you, indirect prompts that only describe the goal, and negative prompts), record which tool actually ran for each, name tools as domain.action, and start descriptions with "Use this when..." including the cases where it does not apply. For example:

Use this when the user wants to search their Acme notes by
title or content. Do not use for reminders or calendar events.

Also mark read-only tools readOnlyHint: true, tools that do not delete or overwrite user data destructiveHint: false, and tools that reach the public internet openWorldHint: true. After launch, review tool-call analytics weekly; a spike in "wrong tool" confirmations usually means the metadata has drifted.

What can a crawler read on chatgpt.com/apps pages?

We tested several addresses with curl on 2026-11-12 (using an ordinary browser User-Agent, which may not match what a real crawler sees):

AddressTest result
chatgpt.com/robots.txt200, 4,573 bytes; Allow lines include /apps, /*/apps and /plugins; CCBot, Google-Extended, anthropic-ai, Claude-Web, PerplexityBot and others are blocked by name
chatgpt.com/apps/ and /plugins200, about 711 KB; the title is just "ChatGPT", with no description, canonical or structured data; only 9 visible characters once scripts are removed
chatgpt.com/apps/spotify/ and /apps/canva/200; title, description and canonical all present, robots is index, follow, with multilingual hreflang; about 5,000 visible characters, with the introduction and example prompts in the HTML
chatgpt.com/apps/zillow/, /figma/, /coursera/200, but every title is "ChatGPT", the app shell; we did not verify whether these apps exist in the directory
chatgpt.com/sitemap.xml200, 80 URLs, none starting with /apps or /plugins

Conclusion: the directory overview pages are empty shells to crawlers; only the introduction pages OpenAI made for a few partner products (we saw Spotify and Canva) carry full content in the HTML. Whether an ordinary developer's directory entry has a public page or gets indexed, we found no evidence; unverified. So do not treat the directory as your only entry point: your own site should have a page saying what the app does and how to connect it, linking to the directory. Whether crawlers run scripts is covered in Do AI crawlers render JavaScript?.

What should a developer do now?

  1. 1. List how users will ask

    The official use-case guide has you list direct and indirect requests first, then decide for each whether a skill or an MCP tool handles it. For example, "Show my Acme launch board" and "What is blocking the launch?".

  2. 2. Write the name, description and example prompts to the rules above

    No comparisons, no prices, and nothing that steers the model to prefer you.

  3. 3. Write a matching page on your own site

    Say what it does, which data it reads and writes, how to connect, and link the privacy policy. Crawlers can read this page; whether they can read the directory entry is unverified.

  4. 4. Ask in ChatGPT yourself, three times

    Do not name your product; describe the user's goal, ask the same question three times, and note whether your tool was called or you were mentioned; see How do you measure AI visibility? Four metrics and a table you can follow. For general recommendation tactics see How do I get ChatGPT to recommend my product?.

FAQ

How are the Apps SDK and plugins related?

The docs site address is still apps-sdk, but the pages now say Plugin and Plugins Directory, and the changelog uses both names. OpenAI never says in one sentence whether this is a rename, so we follow the docs' wording and treat them as one system (inference).

How do users find my app in ChatGPT?

The docs describe the directory and the model deciding when to call it; directory ranking rules are not published; unverified.

What is the keywords field for?

The docs describe it as an optional array of discovery terms and do not say how it feeds directory search; unverified.

Can the description mention price or compare with other products?

No. The guidelines forbid comparing with or disparaging other products and forbid advertising price, subscriptions, discounts or promotions.

What do I need to prepare for review?

Identity verification, a ZIP package, a public HTTPS MCP server and the domain-verification token, a privacy policy, 5 positive and 3 negative test cases, a walkthrough video and release notes; a test account too if sign-in is needed.

Do I need a new review after changing my tools?

The docs say changes to the MCP server are scanned automatically, and eligible updates go live after passing the automated checks; changes to metadata or skills need a new ZIP.

Will directory entries be indexed by search engines?

We found no evidence; unverified. Test on 2026-11-12: the chatgpt.com/apps overview is an empty shell to crawlers, while a few partner pages (Spotify, Canva) carry content in the HTML.

How do I know whether ChatGPT recommends my app?

Do not name the product; describe the user's goal, ask the same question three times, and note whether your tool was called or you were mentioned; the official optimization guide also suggests reviewing tool-call analytics weekly.

Sources

kagentx

Can crawlers read your pages?
Let the evaluation tell you

The evaluation checks whether crawlers can get in and whether the text is served directly, then asks ChatGPT, Perplexity and Claude (with web search on) your buyers' questions and sees whether you are mentioned. The Basic plan is $1, and results usually take 2–3 minutes.