A git server that keeps its objects in object storage

gg serves git over plain HTTP while its objects live in a bucket rather than on a disk. It hosts its own source: everything on this instance is served by the code you can read in it.

How it works

There is no disk to provision and no replica to keep in step. Packs are content-addressed, so a cached object can never be stale and no cache-coherency protocol is needed; a read is a ranged fetch against commodity object storage, and the cost model is round trips rather than bytes held on a volume.

What works today

• Clone, push and fetch over smart HTTP
• Forks that share objects with their parent
• Pull requests with inline review comments
• Web UI, JSON API and a gg CLI
• Orgs, teams, per-repo permissions and personal access tokens
• Unlimited repositories · 1 GB of storage per user during preview · 100 MB per fork

Not yet

• Git LFS
• GitHub and GitLab sync
• bundle-uri fast clone
• Incremental fetch negotiation
• A larger repository ceiling

Throughput, concurrency and uptime figures are published once measured, not before. There is no service level during preview.

See it working

Read gg's own repository — the history, the diffs and the review threads are all served by this instance.

Run your own

gg is Apache-2.0 and self-hostable against any S3-compatible bucket. DEPLOY.md carries the configuration model and a worked deployment.