This video explores how Toss's mini-app platform, Apps in Toss, attracted so many services in a short period and how it finds the apps people actually keep using. Its central point is that, when AI makes apps easy to build, clear user value, comfortable UX, and a sincere commitment to ongoing improvement matter more than the number of services. Toss is opening the door for anyone to launch an app while improving recommendations and placement so good apps can be discovered and grow.
1. Everyday Services Inside Toss
The video begins with two small mini apps with clear value. One is Today's Lucky Bingsu, which tells your fortune after you create a 3D shaved-ice dessert. The other is Lowest-Price Gas Stations, which finds the cheapest gas stations on a map. Their features are simple, but one provides fun and the other utility—both can make a good app.
"Even a small idea can become a good app. What matters is offering users different kinds of value, whether that is fun or practicality."
Kim Seryeon, the PO for Apps in Toss, joined Toss in 2021 after working in startups, entertainment, and consulting. As the project's first PO, she has been building development tools for mini apps and the platform that makes services work well inside Toss. Product Design Director Kim Jiyoon joined after working as a UI designer at Samsung Electronics and a product manager at 29CM, expanding her scope from brand design to product design.
Simply put, Apps in Toss is an app store created by Toss. For service creators, it is a place to launch their services. For Toss users, it is a place to access nonfinancial services inside Toss without installing a separate app.
Toss built the platform to move beyond a financial super app and become an everyday-life super app. Even if Toss is good at financial services, building every daily-life service itself is difficult and time-consuming. It therefore invited makers already doing good work in their own fields into a playground where anyone can experiment and launch.
About a year and a half after launch, Apps in Toss had more than 15,000 live mini apps. The video's title says 18,000, but the conversation refers to over 15,000 live apps. Not every app is searchable at once. Toss personalizes the mini apps it shows users in its fifth tab, known as O-tab.
"Being good at financial services and being good at building nonfinancial services were different problems. So we decided to create a playground and invite people who were already doing well."
2. Abandoning the App Builder and Riding the Progress of AI Models
Initially, few services were available to launch on Apps in Toss. Offering variety required enough apps, but even growing from dozens to hundreds was difficult. The team considered why makers would build for Toss rather than another platform. It worked to shorten review and deployment times so ideas could launch faster, and to make SDKs and development tools easier to use.
A turning point came with Claude's Artifacts feature. Vibe-coding tools such as Replit were attracting attention, and many people felt base models alone could not produce satisfactory results. Seeing Artifacts create polished outputs without a separate tool gave the team hope: if people could build services more easily, thousands of apps might arrive.
But it also brought despair. The team had spent months developing an internal app builder to simplify mini-app creation, then canceled its launch after seeing Claude Opus's output. It decided that riding the pace of model improvement was better than competing directly with large models.
"We decided not to compete with the models, but to ride the speed of their growth."
The team began redesigning the platform from something easy for people to read into something easy for machines to read and use. Rather than merely writing developer documentation and guides for humans, it structured them carefully so AI could clearly identify what to read and which tasks to use it for. MCP connected the information and development tools needed to build mini apps, while tasks previously performed through an admin interface were reshaped so AI agents could perform them.
This thinking extended to reconsidering UI and admin screens themselves. If people could authorize AI to do the work instead of finding documents and operating screens themselves, platform capabilities also needed to be offered in units AI could use.
"We assumed the future user would be a machine, not a person, and made everything available in a form AI could use."
After the team supplied app-building tools and MCP, incoming mini apps grew from hundreds to thousands. The question changed from "How do we get people to build more?" to "How do we manage all these apps and help people discover good services?"
3. Apps Grew Fast Enough to Celebrate Every Day—Then Operational Worries Followed
At first, the team celebrated milestones such as 1,000 and 5,000 mini apps with cake. Teammates encouraged one another, saying they should have started the business earlier. As growth accelerated, celebrations came closer together.
"At first, we applauded whenever we reached 500 or 1,000 apps. At some point, we were applauding every day."
When thousands of apps began arriving daily, the joy of celebrating diminished and operational concerns grew. Mini apps run under the Toss name and within its ecosystem, so the team had to check that services worked and contained no dangerous or inappropriate content. Filtering violent, sexually explicit, or malicious services in advance was another challenge.
The space available to show apps to users was also limited, no matter how many good apps existed. Tens of thousands of services could not all fit into a small number of placements, so deciding which apps to put where became important.
The team met makers who were consistently working hard on their services and asked what motivated them. One important answer was the experience of meeting their first users. In other app stores or marketplaces, creators have to promote a service on social media or advertise after launch to attract users. On Apps in Toss, they could see real people arrive and respond. Watching people use their services motivated some makers to build two or three more or keep refining an existing app.
Toss responded with marketing tools such as push notifications and detail pages, as well as more places for apps to appear. It also kept improving recommendations: which mini app to show after a user tried another, and which wording and content to use in introducing it. The team says it uses internal AI and machine learning, along with app characteristics and user data, to choose marketing copy and target audiences.
4. Fewer Than 1% Are Good: Launch, Recommendation, and Boosting
As app numbers grew, a reality became clear. There is a half-joking industry saying that 1% of services are good and nobody uses the other 99%. The actual data showed an even smaller share of good apps. Here, a good app means an app people actually use, not one that looks impressive.
"To us, a good app is one users actually use. But that proportion was below 1%."
The team pointed to the rapid growth of clones as AI drove creation costs close to zero. Previously, a trend might produce a few similar apps. Now, dozens can appear as soon as a trend spreads. The video discusses several mini apps built around a particular trend appearing within just a few weeks.
"When creation costs fall, one trend can produce a flood of similar services all at once."
Apps in Toss addresses this through three stages: launch, recommendation, and boosting. Its basic philosophy is to let anyone build and launch. Among launched apps, those meeting conditions around UX standards, actual usage data, and review performance become recommended mini apps. Editors personally use these and select especially good ones as boosted mini apps, which receive better placement. The team stresses that boosting is a free opportunity, not paid advertising.
No single metric defines a good app. An app that provides enjoyment once, one used monthly, and a travel app used only a few times a year have different usage frequencies. Revisit rates cannot be applied identically to every service. Apps in Toss divides services into roughly 300 categories and evaluates different metrics for each. Time spent may matter for games, visitor counts for some services, and purchase conversion for others.
Ultimately, what matters is whether a service offers clear value and whether users continue to appreciate it. Today's Lucky Bingsu provides the fun of making shaved ice and checking a fortune; Lowest-Price Gas Stations provides the utility of easily finding cheap fuel. An app can be good without complex features if its value is clear.
5. UX Principles for Comfortable Mini Apps
Apps in Toss has published around five UX principles for recommended mini apps in its developer center and blog. These are less about grand design theory than about ensuring users do not feel uncomfortable or confused when a service behaves unexpectedly.
The first principle is to minimize login and notification-consent requests. Demanding login immediately can drive users away before they even discover the service. The team says its metrics showed that forcing login at the start caused roughly half of users to leave.
Mini apps inside Toss can use an anonymous identifier, or "unknown key", to distinguish users. This supports basic user identification and push notifications without login, so the recommendation is to request login or personal information only when genuinely needed. Tools also let users consent to conveniently share information already stored in Toss, such as a shipping address.
"If you demand login the moment someone enters a service, they may leave immediately. It is better not to force login unless it is essential."
The second principle is to tell users in advance when a full-screen ad will appear and avoid obstructing the core action. More ads may seem to increase immediate revenue, but users leave if they encounter advertising before understanding the service's value. Fewer users also means fewer people seeing ads, which may hurt revenue over time.
This does not mean removing all ads. For example, a photo app can offer free photo saving and explain that users can save a video by watching an ad. This gives the ad a natural context. Similarly, a cheap-gas app can show a banner after users select a station, rather than blocking prices on the first screen, so they see the essential information first.
"It is important to show ads at a moment that fits the context, without blocking the core action."
The team admits it is still searching for the right timing for high-revenue formats such as full-screen ads. For now, it gives makers data on ad impressions, user responses, CPM, and related measures. Longer term, it wants advertising intelligence that decides which ad to show, to whom, and when, based on the service and usage context.
The third principle is to provide a consistent experience. Consistency is not just about attractive screens. Button colors and behaviors should not change abruptly between screens, and navigation and back actions should work as expected. AI-designed screens may follow average patterns while missing context, or cram too many features onto one screen. Apps in Toss says it is preparing to integrate a design agent into MCP to reduce these problems.
The remaining principles center on matching users' expectations with actual outcomes. If tapping "Receive" unexpectedly opens an ad, users may not understand why. A button labeled "Watch an ad to receive points" tells them what they must do and what they will get. Toss tested clearer button copy in its own points service and found that disclosing the ad in advance created a more predictable experience.
"Good UX does not have to be flashy or flow effortlessly. It means showing users the result they expected."
6. Makers Who Survive the AI Era—and the Increasing Scarcity of Sincerity
People who built good services on Apps in Toss were not necessarily experts in the relevant field. Often, they noticed an inconvenience in daily life and wanted to solve it. People from very different backgrounds, including teachers, turned problems they experienced into services.
The key is not stopping after building and uploading an app once. Makers who meet real users, examine data, improve the service, and release it again survive longer. Plenty of services on Apps in Toss are launched and abandoned, while relatively few makers keep improving them in response to users.
"To make a service succeed, you have to meet users and keep improving it. In the end, the person who stays to the end wins."
There is anxiety that AI models will quickly catch up even with the quality and creativity of human work. Kim Jiyoon suggests sincerity may become scarcer instead. As AI-generated services and content multiply, the result reveals whether the maker actually enjoyed creating it or sincerely cared about the problem. That care can come through in an app made because its creator enjoyed using it or truly believed the problem needed solving.
"The sincerity of having built something because you truly wanted to solve a problem will become increasingly rare."
The conversation anticipates people using small, varied services that solve specific problems in the moment, rather than always relying on a handful of major services. Beyond users searching for themselves, AI or machine learning may recommend contextually relevant mini apps and connect people with the services they need.
It also discusses Chinese mini-program platforms such as Alipay and WeChat. If someone orders an Americano every evening, for example, the platform might ask, "Would you like your usual Americano?" A single tap could complete both ordering and payment. Connecting small services within a larger platform in this way lets people get things done more efficiently within limited time.
7. A Playground for Makers—and What Comes Next for Apps in Toss
As a financial-services company, Toss says it often creates a small sandbox first when experimenting with a new technology or paradigm. It provides a safe space to try things before all the rules and infrastructure are in place, then develops procedures and policies based on the results.
In a world changing rapidly through AI, Apps in Toss likewise wants to be a useful testing ground for makers: a platform that helps them quickly implement ideas and see how real users respond.
Asked about the benefits of building on Toss, the team replies:
"Just build a good service, and we will bring you the users. We want you not to have to market it separately."
It adds that, once people begin using services properly, it is still considering how to turn the value of the platform's placement space into appropriate revenue. For now, the direction is for makers to focus on good services while Toss supports user acquisition and fulfills the platform's role.
Closing the conversation, the host says they would prefer to establish principles and build one app well rather than churn out many. Apps in Toss's idea of a good service is similarly about value people actually use and appreciate, rather than simply increasing the count.
Closing Thoughts
AI dramatically lowers the cost and time required to build apps, but also causes similar services to flood the market. Future competitiveness therefore comes less from how many you build than from understanding users' problems accurately and continuing to improve. Apps in Toss's standard for a good app is clear: it offers users value, is comfortable to use, and has a maker who takes responsibility for refining it to the end.