Meta has soft-launched a standalone app called Pocket that converts plain-English text prompts into playable mobile experiences called gizmos — and, per Meta's own Help Center, feeds every interaction back into its AI training pipeline. The app went live on the Apple App Store and Google Play on June 29, 2026, with no press release and no official announcement, and remains available only in Brazil. That silence looks deliberate: on July 7, four US states disclosed that they are collectively seeking $1.4 trillion in damages from Meta over social media addiction harms to minors, the same day Gizmodo and Engadget broke the scope of the lawsuit. Meta's newest product, aimed squarely at the casual-game-creation audience, arrived in that legal context and has provided no public details on how it plans to moderate the AI-generated experiences its users will build and share.

The app's core mechanic is straightforward. A user types what they want — "a drawing game where a blooming flower is the paintbrush," to use Meta's own example — and Pocket's AI generates a playable result in seconds. Meta calls each creation a gizmo, defined in its Help Center as "an interactive, playable AI-generated experience." Gizmos can respond to touch, phone tilts, swipes, shaking, sound effects, and music. Some can access the device camera or pull photos from a user's camera roll. Some, Meta says, can "reason about the world around them." No programming knowledge is required at any step.

What happens to those interactions once a user starts playing matters as much as the mechanic that generates them. Meta's Help Center states plainly: "Your interactions with gizmos on Pocket will be used to improve AI at Meta." A regional qualifier follows: depending on where a user lives, interactions may also be used to personalize content and ads. If Meta's AI cannot answer a prompt, the text a user sends may be shared with unnamed "select partners." That data pipeline is not a side note — it is the economic model. Pocket is free, and the gizmo creation workflow produces a continuous stream of labeled behavioral data: which prompts generate which mechanics, which mechanics hold attention, which sensor combinations produce the most play time. Every tap and tilt trains the next iteration of Meta's AI.

How a Text Prompt Becomes a Playable Gizmo

Pocket's technical pipeline separates it from traditional no-code game tools and explains both its potential and its risk profile. When a user submits a prompt, Meta's AI — almost certainly a fine-tuned variant of the model family powering Meta Superintelligence Labs, the unit that launched Muse Spark, Meta's first proprietary frontier model, in April 2026 — interprets the natural-language description and generates code describing an interactive experience. That generated code must then execute somewhere. This is the point where Pocket's architecture intersects directly with Apple's App Store enforcement actions from earlier this year.

In March 2026, Apple blocked App Store updates for vibe coding platforms Replit and Vibecode, then removed a third app, Anything, from the store entirely. The cited rule in each case was Guideline 2.5.2, which states that apps "may not download, install, or execute code which introduces or changes features or functionality." Apple's position is that when a vibe coding app generates a new experience and runs it inside an in-app web view, the host app is effectively executing downloaded code that changes its own functionality after review — which the guideline prohibits. Pocket generates gizmos from text prompts and displays them as playable experiences inside the app. Whether Meta has sandboxed those experiences in a way that satisfies Apple's interpretation of 2.5.2 — or whether it has an agreement with Apple that the other platforms lacked — has not been disclosed. Pocket is currently listed on the iOS App Store and available to Brazilian users. What happens when Meta attempts a broader rollout depends on that unresolved architectural question.

The precedent is uncomfortable for anyone tracking Pocket's expansion timeline. When Anything tried a browser-based workaround — routing generated app previews to an external browser instead of running them in-app — Apple blocked that version too, then removed the app entirely. A security researcher at Next Reality noted that moderating a gizmo feed is structurally harder than moderating photos or videos, because the question is not whether content is harmful but whether an experience does something harmful — misusing camera access, extracting data it should not reach, or using the remix mechanic to distribute a malicious behavior inside something that looks benign.

Pocket Came From Gizmo. Gizmo Came From Snapchat Veterans.

Pocket did not emerge from Meta's internal skunkworks. The app is the direct product of Meta's acquisition, earlier in 2026, of the team behind Atma Sciences, a startup founded by ex-Snapchat developers that had already built and shipped a nearly identical platform under the name Gizmo. Gizmo accumulated 635,000 lifetime installs across iOS and Google Play and a 98% positive sentiment rating before Meta arrived, according to app intelligence firm Appfigures. Meta secured a non-exclusive license to Atma Sciences' underlying technology as part of the deal. The similarities between Pocket and the original Gizmo app are deep enough that Pocket's Android package ID still reads as com.facebook.gizmo.

The Gizmo team reports into Meta Superintelligence Labs, the same unit behind Muse Spark. That organizational alignment matters for understanding what Pocket is for. One analyst at FourWeekMBA framed it as follows: Pocket is not primarily a game platform; it is a proof of concept for a new kind of vertical integration, moving from AI-assisted development to owned distribution to proprietary behavioral data to advertising revenue. Meta can generate interaction data internally through its own apps at near-zero marginal cost, which makes external data partnerships less necessary and gives the company a new training signal it did not previously own: the full loop of creative intent (the prompt), generative output (the gizmo), and engagement behavior (how users actually play it).

What AI Game Creation Actually Is — and Where It Has Already Broken

Vibe coding — the term coined by Andrej Karpathy, an OpenAI co-founder, in a February 2025 post and named Collins Dictionary's Word of the Year for 2025 — describes software development in which a large language model writes code from natural-language descriptions, often without the human reviewing what was generated. The approach democratizes software creation; it also introduces documented security risks that scale directly to consumer platforms.

A December 2025 analysis by CodeRabbit of 470 open-source GitHub pull requests found that AI co-authored code contained approximately 2.74 times more security vulnerabilities than human-written code. A separate scan of 127 AI-generated apps conducted between March and June 2026 by a firm called Launch Ready Code found that 71% had at least one vulnerability serious enough to enable a data breach, account takeover, or regulatory incident. Only 12% of scanned apps met a production-ready security threshold, and every app in that minority had undergone manual security review before deployment. Security researchers at Wiz found at least one vibe-coded app's production database left entirely exposed. Researchers at Red Access identified roughly 5,000 publicly accessible vibe-coded apps with no authentication at all, and approximately 2,000 that appeared to be leaking sensitive data including medical records and financial information, findings The Verge covered in June 2026.

None of this is specific to Pocket. Meta may have built sandboxing and permission controls that the category average does not reflect. But as of publication, Meta has provided no public details on how gizmos are reviewed before appearing in users' feeds, what governs which sensor access a gizmo can request, or how a harmful remix would be identified and removed before it spreads through the discovery feed.

The Competition Pocket Is Entering Is Already Heating Up

Pocket is not the only app trying to build a social feed of AI-generated interactive experiences. Sekai, a startup founded by Lucky Zhang — whose previous companies were acquired by Apple and ByteDance — raised a $20 million Series A in June 2026, backed by Khosla Ventures and Andreessen Horowitz's early-stage fund, among others. Sekai's users have created more than 15 million mini-apps at a rate of 200,000 per day, and the company reports average user sessions exceeding one hour. TikTok has also been testing its own in-feed mini-game format as part of a broader push into interactive content.

The competitive dynamic tilts heavily in Meta's favor on distribution. A notification inside Instagram or Facebook — platforms with a combined user base in the billions — is worth more in launch reach than any App Store featuring. Meta's track record is that it uses cross-app promotion to build audiences for new products quickly. Reverse engineer Alessandro Paluzzi, who first spotted Pocket on X on July 2, noted that Meta is likely to promote it across its existing apps once a wider rollout begins.

What distinguishes Pocket from Roblox, the most obvious comparison, is more significant than the draft coverage of this story has acknowledged. Roblox requires creators to write actual code — specifically, Lua scripts in Roblox Studio. As of the fourth quarter of 2025, Roblox had more than 144 million daily active users, and more than half of them were under 16. Pocket requires no code. The creative barrier is a sentence. That difference in friction is real, and it positions Pocket not as a competitor to Roblox's creator tier — which requires months of skill development — but as a competitor to Roblox's player tier: the vast majority of users who play games others have made but never learn to build their own.

What Meta Does With Your Data on Pocket

Meta's official disclosure language on Pocket is more extensive than the app's low-key launch might suggest. In addition to the AI training use, Meta's Help Center notes that Pocket collects account information including name, username, profile details, age, and account status related to Community Standards and intellectual property violations, using it "to personalise the experience and help keep the platform safe." If Meta's AI cannot fulfill a prompt, the text of that prompt and general location information may be shared with unnamed third-party partners to improve results. Interactions that violate Community Standards are subject to the same enforcement applied to other Meta content — account restriction or disabling — though Meta has not described how it will identify Community Standards violations inside AI-generated interactive experiences, which present moderation challenges distinct from static images or videos.

On the advertising side, the regional qualifier in Meta's Help Center is significant. In jurisdictions where Meta is permitted to use AI product interactions for ad personalization, the gizmo feed represents a new behavioral signal: not just what content a user watches or likes, but what kind of experience they ask AI to create for them, how long they engage with it, and which sensor features they interact with. Creative intent, expressed through natural language, is a richer signal than passive content consumption.

Users in Brazil who encounter Pocket now and users elsewhere who encounter it once it expands can limit some exposure through Meta's standard privacy controls, but the platform's core data collection — training data use from gizmo interactions — is not opt-out; it is the stated default. Meta's Help Center notes that users "can object to their information being used for AI development through Meta's privacy controls," though the specific mechanism for doing so within Pocket is not described in the Help Center documentation available as of publication.

Is This Where Vibe Coding Goes Next?

The vibe coding tools market has reached an estimated $4.7 billion in 2026, growing at a 38% compound annual rate, according to industry estimates compiled by Taskade and Keyhole Software. Apple's App Store received 235,800 new submissions in the first quarter of 2026 alone — an 84% increase year over year and the largest quarterly surge in a decade, driven largely by AI-generated app creation. The flood of vibe-coded submissions has contributed to App Store review times stretching from the usual 24 to 48 hours to as many as 30 days in some cases.

Roblox and Unity shares each declined during the trading session following Pocket's public discovery, a signal that traditional gaming creation platforms read the launch as meaningful competition. Whether the threat is real depends on a question no one can answer from the app's current Brazil-only soft launch: whether AI-generated gizmos can hold user attention as well as crafted games do. Social Media Today noted that the majority of vibe-coded experiences, because their creators lack design experience, tend toward the basic — and if a feed fills with uninspiring gizmos, the scroll habit that makes TikTok and Instagram addictive will not form. The format has to produce enough compelling experiences to sustain a discovery feed.

Pocket's predecessor, the original Gizmo app, reached 635,000 installs and near-perfect user sentiment before Meta acquired the team. That suggests the format has an engaged audience at indie scale. Whether it has one at Meta scale is the experiment Pocket is designed to run.

Frequently Asked Questions

What is a gizmo on Meta Pocket?

A gizmo is Meta's term for an AI-generated interactive experience created inside the Pocket app. Users type a description of what they want — a physics puzzle, a tilt-controlled game, a drawing tool — and Pocket's AI generates a playable result. Gizmos can respond to touch, phone tilt and shake, sound, and camera input. They can be shared in a social feed where others play and remix them. The term comes from Atma Sciences, the startup whose team and technology Meta acquired earlier in 2026.

Will Meta Pocket work on iPhone long-term?

It is currently available on the iOS App Store in Brazil. Whether that availability survives a broader rollout depends on an unresolved question about Apple App Store Guideline 2.5.2, which prohibits apps from executing code that changes their functionality after review. In March and April 2026, Apple blocked Replit and Vibecode from publishing updates under this rule and fully removed Anything from the store. Those apps generated interactive experiences inside in-app web views — the same mechanic Pocket appears to use for gizmos. Meta has not disclosed how Pocket's execution architecture differs, or whether it has reached an agreement with Apple.

Does Meta Pocket use your camera and interactions to train its AI?

Yes. Meta's Help Center states that interactions with gizmos on Pocket are used to improve AI at Meta. Camera and camera roll access are optional and user-controlled at the device level, but Pocket requests those permissions. Gizmo creation prompts, engagement patterns, and play-session data are collected by default. In some regions, this data may also be used to personalize ads. If Meta's AI cannot fulfill a prompt, the prompt text may be shared with unnamed third-party partners. Users can object to AI training use through Meta's privacy controls, though the specific mechanism is not documented in the Help Center materials available as of publication.

What makes Pocket different from Roblox or traditional game apps?

Roblox requires creators to write actual code — Lua scripts in Roblox Studio — which takes months to learn. Pocket requires a sentence. That difference in friction is the product's entire premise. Traditional mobile games are made by developers using engines like Unity or Unreal. Pocket generates each experience on demand from a text prompt, which means creators have no technical knowledge of what runs on their phone when they play it. That architectural difference is also why the Apple enforcement questions matter: a traditionally developed app runs reviewed, static code; a Pocket gizmo executes generated code that did not exist at review time.