The cycle is familiar to anyone who's worked at a software company. A product manager spots a typo on the homepage. Marketing wants to test a new call-to-action button. A designer needs to adjust padding on the mobile view. Each request—minor on its own—gets logged, prioritized, and typically lands behind a dozen other engineering tasks. Weeks evaporate. The original requester follows up. Eventually the change ships, sometimes months after it mattered.
Daniil Bekirov thinks that's absurd. So the 20-year-old built Sparkles, an AI code editor meant to eliminate the wait entirely.
Bekirov dropped out of University College London to join Y Combinator's Winter 2026 batch—a choice that perhaps signals confidence more than caution. His pitch is deceptively simple: let anyone on a team make changes to a codebase without breaking production. "Make everyone on your team an engineer (without breaking anything on prod)," the company promises, a tagline that somehow manages to be both ambitious and hedged.
Sandboxes with training wheels
Here's how it works. Sparkles spins up isolated development environments tied to a company's existing GitHub repository. A technical owner—still a developer, for now—connects the repo and hands over environment variables. After that, anyone with a company email can log in, describe what they want in plain language, and watch the AI make edits in a live preview environment.
Like what you see? Push it as a pull request back to GitHub. An engineer still reviews it, of course. The guard rails remain. But the initial heavy lifting happens elsewhere.
The tooling borrows liberally from developer workflows while stripping away everything that typically requires a computer science degree. Git commands? Gone. Local environment setup? Irrelevant. Sparkles claims to support major frameworks—Next.js, Vite, others—and says it can import repositories in a single click. The AI runs on Anthropic's Claude, with Cloudflare infrastructure handling the sandbox isolation. Whether that setup can handle edge cases at scale is the kind of question that typically gets answered in production, not pitch decks.
The comparison Sparkles makes most often is to Lovable, another AI code editor that's gained traction recently. But Lovable's documentation suggests it doesn't yet support directly importing existing GitHub repositories as new projects—a limitation that Sparkles is betting matters. "Like Lovable for existing projects," the positioning goes, targeting the messy reality most companies actually inhabit: codebases already in flight, technical debt already accrued, teams already stretched thin.
The economics of delegation
Pricing follows a familiar SaaS ladder. Free users get 100 credits monthly and one project—enough to kick the tires. Pro runs $20 for 500 credits and five projects. Ultra, at $80, unlocks 1,000 credits and unlimited projects. Teams can opt for a "Pilot" plan with GitHub integration and priority support; enterprise customers get on-premises deployment and full data control, the table stakes for selling to anyone with a security team.
Bekirov has reportedly promised personal onboarding for early adopters, according to LinkedIn posts around the launch. Demo slots filled quickly, which may indicate genuine demand or simply reflect the small scale of an initial rollout. Hard to say yet.
A suddenly crowded space
Sparkles isn't alone in sensing opportunity here. Vercel's v0—a tool that's already won over designers who want to prototype quickly—recently added Git Import functionality that lets non-engineers open pull requests against existing repos. Builder.io launched Fusion, which connects design systems to repositories and uses an AI bot to commit changes straight from PR comments. CodeSandbox and StackBlitz have offered cloud-based development environments with PR workflows for years, though they've largely marketed to developers rather than cross-functional teams.
What Sparkles is doing differently, or at least claiming to, is democratizing access. This isn't a tool for engineers who want AI assistance—it's infrastructure meant to extend coding capability across entire departments. The final review still belongs to engineering, but the initial creation doesn't. That matters if you believe the real cost isn't the five minutes it takes a developer to change a button color, but the context-switching penalty every time they're pulled away from deeper work.
GitHub seems to agree there's something here. Last year the company announced Copilot's expansion to multi-model support and introduced "Spark," a feature for natural-language app creation. Google followed in December 2024 with Jules, an AI agent that fixes bugs and creates pull requests autonomously. The industry pattern is clear: AI-generated code changes flowing through standard review processes are becoming normal, maybe even expected.
Still early days
The fine print reveals how nascent this is. Sparkles operates under the legal entity Litmus Labs, Inc., with Terms of Service dated January 1, 2025—barely two months old as of publication. The homepage nudges visitors to subscribe for updates and upvote on Launch YC, the kind of call-to-action that screams staged rollout rather than general availability. Y Combinator lists Bekirov as the sole team member, with Jared Friedman noted as the batch's primary partner. Solo founders at 20 either move fast or burn out; not much middle ground.
The company's privacy policy does address one anxiety that comes up immediately with AI coding tools: Sparkles promises not to train models on customer code without explicit consent. That's reassuring, though it's also the kind of thing that's easy to promise when you're small and gets complicated when you're trying to improve your product at scale. The vendor stack includes Clerk for authentication and Convex for database services—both solid choices, nothing exotic.
The question that lingers
Whether sandboxed AI editing becomes standard practice or remains a niche workflow for specific teams probably depends less on technology than on culture. Some engineering organizations will see this as a reasonable way to reduce bottlenecks. Others will worry—legitimately—that democratizing code changes invites chaos, even with PR review as a safety net.
Sparkles is betting that the pull request, an existing and trusted mechanism, provides enough oversight to make the risk acceptable. For teams that buy that logic, the value proposition is hard to argue with: fewer bottlenecks, faster iteration, and a button color that gets changed this week instead of next quarter.
Then again, the button color might not have needed changing at all. Sometimes the bottleneck is the point—a filter that separates signal from noise. Bekirov's gamble is that most companies have too much friction, not too little. He's probably right about that. Whether his solution works as advertised is the kind of thing that only real-world usage will reveal.
For now, the pitch is compelling enough that people are lining up to try it. That counts for something.
