When GitHub Copilot launched and began auto-completing lines of code – and, later, entire code snippets – the question many people were asking was: how long until we can simply describe an application in natural language and that Copilot builds it for us? We’ve seen a lot of experiments in this area in recent months, but now GitHub itself is backing this idea with the announcement of GitHub Spark at the company’s annual GitHub Universe conference in San Francisco.

Spark, which is officially an experiment the company is launching from its GitHub Next Labs, lets you quickly create a small web application using only natural language. Experienced developers can still see and edit the code – and underneath that is a GitHub repository, GitHub Actions, and Microsoft’s Azure CosmosDB as the default database for apps that need it – but it’s optional. Ideally, you’ll be able to use a chat-style experiment to create a prototype and then refine it in subsequent steps.

To create an app, you start with an initial prompt and Spark will show you a live preview in a few seconds. From there, you can refine the app and iterate by going back and forth with the bot.

“We truly believe this is the next step in building personal software,” GitHub CEO Thomas Dohmke told me. “These micro-applications are not intended to replace the professional developer. They are intended to be a tool for you to explore ideas, create little robots, little assistants in your daily life – or simply explore software development with natural language.

While the GitHub docs talk about Spark as a tool for building “micro-applications,” Dohmke told me he’s not really sure if there are any limitations built into the complexity of a Spark app and a GitHub spokesperson confirmed that there was indeed one. no limits on prompts or capacity at this time (although this may change after the technical preview). Dohmke noted that Spark can use any web API and leverage AI models itself and that he’s excited to see how far GitHub users will push this tool.

“It will be very exciting to see, honestly, what users can do, how big of an application they can create with just the power of natural language and this simple user interface that is really designed to let you explore ideas, to generate new ideas. “, he said.

Given that GitHub Copilot now allows users to choose which large language model they want to use, it’s no surprise that Spark offers the same functionality, with users having the choice between recent versions of Anthropic’s Claude Sonnet and OpenAI’s GPT models.

Users can easily share their Sparks with personalized access controls. Perhaps even more interesting is that these users can also take the shared code and develop it themselves.

Developers who want to take these applications even further can review the code at any time and modify it if there are any problems, which Dohmke acknowledges could arise. AI isn’t perfect, after all. “You can of course look at the code base,” he said. “So if you understand the code base, you can also look at the code directly and edit it, which is often helpful when the AI ​​makes a mistake, which it happens.”