Original research · August 2026
We catalogued every app in the public ChatGPT directory: 2,305 listings, 13 categories, 2,007 companies. Not a sample. Here is what the full population shows about who is building, what they are building, and where the gaps are.
Method
In August 2026 we walked every category in the public ChatGPT app directory through the directory’s own listing interface, following pagination to the end of each category rather than reading the handful of cards on the landing page. That produced 2,305 listings.
For each listing we captured category membership, developer, description, website, published policy links, and authentication policy. 16 listings were not installable at the time of capture and are excluded from the figures below, leaving 2,289 apps. This is the full listed population, not a sample, so the counts have no sampling error. They do have a shelf life: the directory changes continuously and these figures describe August 2026.
One measure we deliberately do not report is launch volume over time. The directory exposes a creation timestamp, but the overwhelming majority of records share a single narrow window, which is characteristic of a platform migration rather than of when the apps were actually built. Charting it would look impressive and mean nothing.
Finding 01
The five largest categories hold 73% of everything listed. The five smallest hold a rounding error, despite mapping to enormous software markets.
Business & Operations holds 440 apps. Security holds 19. Security software, communications platforms, and analytics vendors are all industries worth tens of billions of dollars with fewer listed apps between them than a single mid-size category.
Findings 02–04
1,918 of 2,007 developers, or 95.6%, have exactly one app listed. The most prolific third-party developer has 29 apps. In a mature app store, a long tail of single-app publishers sits underneath a head of established multi-app companies. Here there is essentially no head.
Only 154 apps, 6.7% of the directory, define named skills: the structured capabilities that let an assistant carry out a task rather than answer a question. 70% ship even a single example prompt.
This is the most actionable finding here. The gap between a thin read-only listing and a connector that actually does work is where the competitive difference sits right now, and it is a much lower bar than most companies assume.
2,260 apps, 98.7%, publish both a privacy policy and terms of service. 99.7% authenticate at install rather than deferring it to first use. 96.7% link to a real product website.
Rates that high across a population this size are not a coincidence of good practice. They describe what passes review. A company planning its first connector can treat published policies, install-time authentication, and a real product presence as prerequisites.
Where the gaps are
Ranked by how little is listed against them. Each links to every app currently in that category.
One of the smallest buckets, and conspicuously conservative: most listings describe themselves as read-only and route anything consequential back to the vendor’s authenticated console, which is the right pattern for the category.
Messaging, email, telephony, meetings, and contact-centre platforms. The recurring shape is read-then-send: search over threads and history is near-universal, sending is present but scoped narrowly, behind explicit confirmation.
Small relative to the size of the sector, which reflects how carefully the category has to be entered. Listings avoid diagnosis and stay on information, navigation, and administrative tasks, with narrow, carefully scoped retrieval surfaces.
Generation and editing utilities dominate. Because the assistant can already produce text and images, listings here compete on the specific asset library, brand system, or export format they control.
Warehouses, BI platforms, semantic layers, and a long tail of dataset-specific lookup services, mostly built by the platform vendors themselves. Two shapes recur: a small surface wrapping one query primitive, and a wide surface exposing every object in a warehouse.
What it means
The headline is that this market has not consolidated. A directory where 95% of participants have shipped one app, where fewer than 7% define real capabilities, and where several major software categories are close to empty is not a market where being late is fatal. It is one where being deliberate still counts for more than being early.
The second point is about depth. Presence in the directory is not the differentiator it might have been a year ago, because presence is cheap. What is still rare is a connector with a well-designed tool set, a real permission model, and enough scope to answer the questions customers actually ask. Most of what is listed does not clear that bar.
The third is about portability. Everything in this directory is built on the Model Context Protocol, which is the same protocol behind Claude connectors and Microsoft Copilot’s agent surfaces. The engineering does not have to be spent per platform, and treating a ChatGPT listing as the goal rather than as the first distribution channel is the most common strategic error we see.
Reuse
The compilation, categorisation, and analysis on this page are published under CC BY 4.0: free to quote, republish, and build on, including commercially, with attribution to Node8 and a link to this page. The underlying facts come from OpenAI’s public directory and are not ours to license. Please include the snapshot date, August 2026, since the directory changes continuously.
Suggested citation: Node8, The State of ChatGPT Apps, August 2026. https://node8.ai/state-of-chatgpt-apps
Journalists, analysts, and researchers who want the aggregate figures in a machine-readable form, or a cut we have not published, are welcome to get in touch.
FAQ
2,305 listings were present in the public directory when we captured it in August 2026. 16 were not installable at the time, leaving 2,289 available apps across 13 categories.
2,007 distinct developers are named across the directory. 95.6% of them have exactly one app listed, which is the clearest single indication of how early this market is.
34 of 2,289, or about 1.5%. The other 2,255 come from third parties. That split is what a marketplace is designed to produce rather than a finding about it, so we would not read anything into the ratio itself.
Published policies and a clear authentication model, at minimum. 98.7% of listed apps publish both a privacy policy and terms of service, and 99.7% authenticate at install rather than deferring it. Those rates are high enough and consistent enough to treat as requirements rather than as common practice.
Security (19), Communication (24), Healthcare (53), Creativity (88), Data & Analytics (89). Each maps to a substantial software market with very little listed against it, in a directory where Business & Operations alone holds 440.
Yes. The compilation, categorisation, and analysis are published under CC BY 4.0, so you may quote, republish, and build on them, including commercially, provided you attribute Node8 and link to this page. The underlying counts come from OpenAI’s public directory and are not ours to license. If you want the aggregate figures in a machine-readable form for research or reporting, get in touch.
Every category in the public directory was walked through the directory's own listing interface in August 2026, capturing each app's category, developer, description, policy links, and authentication policy. No sampling: this is the full listed population as of that date. Figures will drift as the directory changes.
Node8 builds MCP servers and ChatGPT apps end to end: tool design, authentication, platform review, and the adoption work afterwards. The same server serves Claude and Copilot.