A Mac mini in our fleet became unresponsive after 37 days of uptime. The kernel had wired 15.1 GB of the machine’s 16 GB of memory. Only 67 MB remained free, the load average was 250, and the system could no longer fork a process. Photon runs Mac hardware continuously for our iMessage API relays. We found that every Mac we checked was accumulating wired memory at a steady rate. At the observed rates, the fastest machine would exhaust its memory in about five weeks. The cause was a kernel memory leak triggered by nested `#!` scripts. The leak can be reproduced with six lines of shell. It does not require administrative privileges, and the allocated memory is released only after a reboot.

NanoClaw now offers hosted iMessage support through Photon, eliminating the need for a local Mac, personal Apple ID, or hardware management; agents can send and receive texts, attachments, reactions, typing indicators, read receipts, and effects on any OS via Photon’s cloud infrastructure, with simple setup using a Photon account and NanoClaw’s setup commands.

Photon processes over 10 million iMessage API calls per day across production deployments for companies like Rho, Vercel, Hermes Agent, Ditto, and FlipText. The platform achieves the same uptime, delivery speed, and feature coverage as Apple's own iMessage service — not through incremental improvements to the standard Mac-bridge architecture, but through a fundamentally different technical approach. This post breaks down the three engineering layers that make that possible: protocol-native Apple communication, Astrolabe (our open-source macOS fleet manager), and enterprise cloud routing on AWS EKS.

Your iMessage lines now scale themselves. Route users to the healthiest line with a single API call.

Spectrum now supports live, interactive mini app cards inside the iMessage transcript — plus in-place updates so cards can react in real time.

We upgraded our Chat SDK iMessage adapter to Spectrum — the result is the most feature-rich and reliable iMessage integration in the ecosystem, with tapbacks, effects, mini-app cards, webhooks, and more.

We migrated the iMessage routing service from Bun to Node, added OpenTelemetry tracing, replaced a fragile fan‑out architecture with a durable Postgres‑backed event log, fixed a binding‑cache bug that dropped messages on transient errors, and switched the OTLP exporter to gRPC, resulting in a simpler, observable, and more reliable system that can handle over a million messages daily.