In early 2025, 37signals transitioned their container registry from external providers Dockerhub and Amazon ECR to an on-premise solution using Harbor, supporting their cloud exit and deployment with Kamal. Challenges with external registries included high costs due to bandwidth overages, performance issues causing delayed deployments (notably a 45-second penalty on uncached pulls), security concerns over potential credential leaks, and loss of operational independence due to API limitations. They selected Harbor for its rich features, scalability, reliability, and open-source nature. The initial setup involved two standalone Harbor instances deployed with docker-compose, one in Ashburn and one in Chicago, configured with their own S3-compatible Pure FlashBlade storage cluster. Minimal necessary S3 permissions were carefully applied for security. Replication between the two Harbor sites was implemented as a two-way synchronization every 10 minutes, using Terraform’s Harbor provider to configure replication policies. Underlying S3 buckets were replicated separately but this alone was insufficient without Harbor-level replication configured. Migrating the existing image catalog from Dockerhub required handling API throttling by creating many individual replication rules per repository, triggered in batch scripts to avoid 429 errors. Scripts were provided for generating Terraform config files from Dockerhub repositories, triggering replication batches, and monitoring replication progress. Performance analysis after migration showed significant improvements: image pull times decreased by up to 25 seconds for their largest applications (HEY, Basecamp 4 and 2), deployment times reduced by 15 seconds for HEY, and storage usage dropped from nearly 9 TiB to 1.5 TiB due to retention policies and garbage collection. They also retired their Dockerhub cache infrastructure and saved approximately $5,000 annually on subscription fees. The single-node primary Harbor endpoint in Chicago with Ashburn as backup has been stable over two months, serving over 32,000 pulls during everyday operations. This project demonstrated the advantages of transitioning away from large SaaS and public cloud providers: improved performance, cost savings, and simplified infrastructure.