Claude App Ecosystem: A Statistical Overview#
As of early 2026 the Claude app ecosystem has grown to encompass tens of thousands of publicly available projects across GitHub repositories, HuggingFace spaces, personal portfolios and enterprise deployments. Analysis of these projects reveals a stark and uncomfortable truth: approximately 95% of all Claude-powered applications fall into the category colloquially known as “AI slop,” derivative and low-effort software that adds minimal value to the world and minimal skill to its creator. Only a slim 5% of Claude apps demonstrate genuine engineering depth, creative ambition or substantive contribution to the broader knowledge commons.
The raw numbers are staggering. Estimates place the total count of Claude-integrated projects at somewhere between 80,000 and 120,000 distinct repositories as of Q1 2026, up from roughly 12,000 at the start of 2024. That growth rate, nearly 10x in two years, sounds impressive until you examine the composition. Of those 80,000 to 120,000 projects, roughly 76,000 to 114,000 are simple wrappers: a few dozen lines of Python calling the Anthropic API, a prompt hardcoded into a string literal and a README that was itself generated by Claude. The barrier to entry has collapsed so completely that the signal-to-noise ratio across the ecosystem approaches zero.
The median Claude app receives fewer than 3 GitHub stars, has a commit history spanning fewer than 7 days and is never updated after its initial push. Approximately 68% of all Claude repositories have zero open issues; not because they are bug-free but because no one is using them. Around 41% contain a single file. About 29% contain only a README and a requirements.txt with no actual implementation code whatsoever. These are placeholders for ambitions that were never acted upon or demos built for a social media post and then abandoned.
The AI slop category is not monolithic. It breaks down into several recognizable subcategories with their own internal statistics.
The largest subcategory, representing roughly 34% of all Claude apps, is the “chatbot wrapper.” These applications take the Claude API and present a custom chat interface with a system prompt that defines some narrow persona: a customer service agent for an imaginary company, a cooking assistant, a fitness coach or a “no-filter” Claude variant. The system prompts average 47 words in length. Roughly 83% of these projects copy their system prompt directly from a tutorial, a Reddit post or another open-source repository with minimal modification.
The second largest subcategory at approximately 22% is the “document summarizer.” These apps accept a PDF, a URL or pasted text and return a Claude-generated summary. They offer no novel chunking strategy, no retrieval-augmented generation, no evaluation harness and no consideration of hallucination rates. They are implemented in an average of 94 lines of code and their authors typically declare them “production ready” in the README.
A further 18% of Claude apps are “automation scripts,” often single-file Python programs that use Claude to rename files, draft emails or reformat spreadsheets. While occasionally useful in isolation, these scripts are published without tests, without error handling for API failures and without any mechanism for the user to inspect or override the model’s decisions. They are published as if novelty alone constituted value.
The remaining 21% of the slop category is a long tail: Claude-powered browser extensions that inject summaries into every webpage; Claude-integrated Discord bots that respond to every message whether asked or not; half-implemented agent frameworks that duplicate LangChain with fewer features and no documentation; and portfolio projects included in job applications by candidates who have never written a line of software outside of a prompt.
The 5% that escapes the slop designation is defined not by complexity for its own sake but by intentionality. These projects demonstrate evidence that their authors read documentation carefully, made deliberate architectural decisions, tested their assumptions and considered the experience of future users or contributors. They are disproportionately represented in areas where domain expertise compounds the value of language model integration: scientific computing, developer tooling, education and systems programming.
One consequence of the 95/5 split is that discovery becomes nearly impossible through conventional means. GitHub search returns slop by default because slop is more numerous. Star counts are gamed, trending algorithms reward novelty over quality and the projects that would most benefit an intermediate developer are buried beneath thousands of chatbot wrappers.
Surveys of developers who use Claude professionally suggest that roughly 71% have wasted more than two hours attempting to use an open-source Claude app that turned out to be non-functional, abandoned or misleading in its documentation. About 54% report having simply decided to write their own implementation after failing to find a trustworthy existing one. This duplication of effort, each skilled developer reinventing the same wheel because the existing wheels are untrustworthy, represents an enormous aggregate waste of engineering time.
The 95/5 ratio reflects structural incentives more than individual character. Claude made it possible to ship something in an afternoon that would have taken a week in 2022. The friction reduction is real and valuable; but it also eliminated the natural filter that friction provides. When writing a working program required sustained effort across multiple days the selection pressure for actually having something to say was correspondingly high.
Today the activation energy for publishing a Claude app is lower than the activation energy for deciding not to. The result is a vast sediment layer of projects whose authors had a thought on a Tuesday afternoon, translated it into code with Claude’s assistance and pushed it to GitHub before dinner. There is nothing wrong with any individual decision in that chain. The aggregate effect is an ecosystem that is difficult to navigate and easy to dismiss.
The 5% exists because some builders carry standards across the new environment rather than abandoning them in it. They ask harder questions: Does this work reliably? Does it handle failure modes? Would I use this myself in six months? Can someone else contribute to it? These questions take longer to answer than it takes to push a repo; but they are the questions that separate work worth finding from work that merely exists.