We're making Git hosting more reliable, performant, and scalable. This post trac
- Source
- cursor.com
- Author
- cursor_ai
- Date

Context
Cursor built the Git storage behind Origin, its code hosting service, around a log of every push kept in cloud object storage, rather than around copies of each repository on servers. The blog post explaining it measures that choice against the design most Git hosts use, which it traces to GitHub's Spokes system from about 2013 (Cursor blog). Spokes keeps several full copies of each repository on fast local disks and accepts a push only when a majority of copies agree, using three-phase commit, a protocol in which servers vote on a change before any of them applies it. Cursor's criticism is that every step waits for the slowest server, so adding copies slows pushes. A large monorepo hit by heavy CI traffic needs more than three copies, while the many small, often throwaway repositories that agents create still need three each. Because the disk copies are the record, each one must also be tracked, checked and quickly repaired.
Cursor's system, Continuity, makes a write-ahead log the record instead: an append-only list of changes written to durable storage, here S3-compatible object storage, before anything else happens. Cursor says a push is acknowledged only after it is fully stored in the log, and becomes visible only once its branch update is prepared on a local copy and recorded in the log's index. An atomic compare-and-swap on S3, a write that succeeds only if nothing changed in the meantime, orders pushes, so any server can safely accept one. Local repositories are ordinary Git repositories used as a cache that can be rebuilt from the log. Replicas learn of new pushes through UDP messages that may be lost, so before serving a read each one asks S3 whether the log index has changed. Cursor says that check takes under 10 milliseconds on average, and a replica that is behind catches up first, which is how it claims every read is consistent.
The payoff Cursor reports is scaling in both directions. In its synthetic tests, read throughput grew linearly up to 100 replicas without slowing pushes, and a cluster sustained up to 120 pushes per second on S3 Standard and more than 300 on S3 Express One Zone. A busy monorepo can run on hundreds of replicas, a small repository on one, and an idle repository on none until it is requested again. These are Cursor's own measurements, not independent results. Origin itself was announced in an earlier Cursor post.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
Checking sign-in…
Loading comments…





