GitHub has announced the official retirement of Spark, its experimental tool that let you create micro-apps by describing them in natural language. As of this Tuesday, August 4, the platform no longer accepts new users or allows creating new apps, and the service will shut down permanently on the upcoming August 31. Projects that are already deployed will keep working after that date, but anyone who wants to keep the code will have to export it before the shutdown takes effect.
What Spark was, the lab of apps-by-description
Spark was a space integrated into github.com where anyone could write what they wanted to build, for example “an app that reminds me when to water the plants”, and get a working application within minutes, without writing a single line of code or reviewing it afterwards. It was part of GitHub’s bet on so-called AI-native development, presented alongside Copilot in late 2025, and its purpose was to test to what extent you can “program by talking”.
To give intelligence to the applications, Spark relied on GitHub Models, the platform’s model inference service. Apps could invoke it through a function called llm(), which added AI capabilities to the project without the creator having to manage keys or external providers.
Why it is disappearing: AI agents have taken its place
The official explanation is as simple as it is significant: the very evolution of artificial intelligence has eaten the product. GitHub assures that, since the launch of Spark, models and agentic development tools have advanced very noticeably, and that creators already build and refine this kind of experience with Copilot in the environments where they work every day: VS Code, the Copilot CLI and the GitHub Copilot app.
“We have seen developers increasingly choose these integrated flows, and we are aligning our offering with that demand,” the company explains in its changelog. In practice, the decision reflects the same trend running through the whole industry: isolated chat interfaces for building applications give way to agents that operate inside the editor, the terminal or the repository itself.
GitHub Models had already been retired in July
The announcement is the second blow in a few days. On July 30, GitHub retired GitHub Models, the inference service on which Spark’s llm() function depended. Since that date, calls to llm() no longer work, so applications that used AI lost that capability all at once. Those that did not use it were not affected by this retirement, although now they all share the same fate: the entire platform goes dark at the end of the month.
What you have to do if you were still using Spark
For users who created projects with the tool, GitHub has left clear instructions. Applications already deployed will keep working after August 31, so there is no urgency on the service side. If you want to keep editing your code in the future, you must export it before that date: open the app’s workbench, press the “…” menu and choose “Create repository”, and the project will become a normal GitHub repository.
And if your application uses llm(), you will have to replace the inference with your own provider: you will need your own API key and manage the billing, because GitHub will no longer provide the model or tokens. To check whether your app is affected, just search for the string llm() in the code: if no call appears, no action is needed.
A symptom of where software development is heading
The closure of Spark is minor news in product terms, but it illustrates an underlying trend. The “build by describing” layer does not disappear: it integrates into agentic workflows. What yesterday was a standalone tool, today is one more capability inside the editor you already use. GitHub is not abandoning the idea of programming with natural language; it considers it so obvious that it has fused it with Copilot.
The practical takeaway for the developer is that AI-generated code stays where it is: in the repositories, the editors and the agents that work in them. Spark dies, but the conversation with the machine that made it possible is more alive than ever, integrated into the place where software is made every day.
