Pando

Pando is OpenAI's Git-protocol-compatible backend (largely Rust) for its monorepo. In a November 2, 2026 engineering post, the company explained how engineer + Codex agent traffic (commits up ~1,300% since early 2024) broke conventional Git serving—without abandoning Git workflows.

Pando has served production since April 2026. OpenAI reports roughly 5× / ~80% better median push latency and nearly a petabyte of Git data delivered per day, aiming at horizontally scalable concurrent pushes and read throughput that can saturate client NICs.

Design choices

Typical Git servers serialize branch updates repository-wide and often build a fresh packfile per fetch. Pando instead:

1. Uses per-branch causal consistency so unrelated branch pushes proceed in parallel (readers watching many branches may see slightly skewed tips).
2. Reuses packfiles via mechanisms such as packfile-uri, served from regional caches (described as Express) with >99% hit rates claimed.

On push, packs stream into blob storage after validation; metadata records branch tips and pack dependencies. Trunk is advertised as the usual base so feature-branch packs stay reusable.

Vs custom VCS forks

OpenAI kept Git clients and protocol and is feeding client improvements upstream—contrasting with shops that replace Git entirely. Pando is internal today; leadership has floated sharing it more broadly.

Who should care

• Platform teams whose monorepos and agent/CI fleets saturate hosted Git.
• Infra engineers studying source control in the agent era.

Next steps

1. Read Scaling Git at OpenAI.
2. Benchmark your own global branch locking and pack-build costs.
3. Watch for any public Pando offering or upstream Git patches—there is no public SaaS yet.