How Cursor beat Git’s scalability shortcomings
Cursor, the AI coding assistant company, has detailed how it built its own Git repository service, called Origin, to overcome scalability problems that have long plagued the distributed version control system, particularly as AI coding agents generate ever more code, pull requests and CI runs. In a technical post, Cursor principal systems engineer Vicent Martí, who previously worked at GitHub, explained that Git's design as a graph-based, content-addressable store forces servers to traverse an entire commit history to fulfil requests, a problem that worsens dramatically at scale. Rather than following GitHub's established approach, Cursor built its system, powered by an internal engine called Continuity, on cloud object storage instead.
GitHub's solution, known as Spokes, keeps at least three synchronised copies of every repository on fast NVMe disks, an approach that became an industry standard but grows slower to synchronise as more replicas are added, and struggles with Git's poor tolerance for "eventual consistency". Cursor's alternative uses Amazon S3 as the source of truth, writing pushes into an immutable write-ahead log while simultaneously updating a single local "reference" copy on NVMe disk; other replicas then sync from there as needed. Martí said this design lets the system ingest pushes as fast as the disk allows, since it only needs to synchronise with one local repository rather than a quorum of replicas. A beta of the resulting service, Origin, is now available to Cursor's paid subscribers.
- Cursor built Origin, a Git service using S3 object storage instead of replicas.
- It addresses scaling issues worsened by AI coding agents generating huge PR/CI volumes.
- Unlike GitHub's multi-replica Spokes system, Origin syncs just one local copy.
New here? Start with this
Cursor is a company best known for its AI-powered coding assistant, a tool that helps software developers write and manage code with the help of artificial intelligence. Behind the scenes, all coding work relies on Git, a widely used system that tracks changes to code over time, and platforms such as GitHub that host and serve that code to millions of developers. As AI tools generate ever larger volumes of code, updates and automated checks, the systems that store and deliver that code are coming under growing strain.
The issue at the heart of this story is a long-standing technical limitation in how Git works at large scale, and how companies attempt to work around it. GitHub, the dominant code-hosting platform, addressed this by keeping multiple synchronised copies of each project on fast storage, an approach called Spokes that became the industry norm. Cursor has instead built its own alternative system, arguing that the growth of AI-generated code makes existing approaches harder to sustain.
This matters because the infrastructure underpinning software development is normally invisible to most users, yet it directly affects how quickly and reliably code can be written, tested and shared. As AI coding tools become more widely used, the systems supporting them are having to adapt, and different companies are taking different technical approaches to solve the same underlying problem.