One Mac hub for routing, memory, and local inference

Day-to-day coding kept breaking when every session pointed straight at a raw model server and hoped routing, memory, and context compaction would sort themselves out. So I built Home Base on Apple Silicon: a hub between the coding-agent client and the model process.

The client in this series is OpenGrok. This post is the hub it aims at over Tailscale.

Series order: OpenGrok, then this post, then Flash Next MTP vs 27B+DFlash, then speculative decode / parent+1.

The shape

Layer What it is
Hub ghost128: 128 GB Apple Silicon, reachable over Tailscale
Brain Home Base proxy on :9081: Jev routing, memory, compact endpoint
Serve mlx-serve Flash Next mixed 4/8-bit on :11234
Client OpenGrok on gHost64 → homebase-brain over Tailscale

The model on the hub today is ddalcu/Qwen3.8-Flash-Next-MLX-Serve-mixed-4-8bit (model_max 262k). Before that default, Home Base ran Qwen3.8-27B-4bit with a DFlash2 drafter at --ctx-size 150000. Why I switched, and the tok/s numbers, live in the Flash Next post.

Why a brain in the middle

Pointing the client at raw mlx-serve worked until compaction and fan-out got messy. The brain owns:

  • Jev routing: what stays in context vs what gets dropped
  • Memory: durable state across turns without stuffing the prompt yourself
  • POST /v1/compact: the client hits this when auto_compact_threshold_percent = 40

I deliberately keep per-turn chat_compact = false. Compacting every turn fights the prefix cache. Better to let the brain compact at the threshold and keep hot prefixes reusable. Live Titan sessions showed ~99% hot-cache reuse at ~45k prompt; details in the Flash Next write-up.

How I use it day to day

  1. ghost128 runs the brain (:9081) and Flash Next (:11234).
  2. gHost64 runs OpenGrok aimed at homebase-brain over Tailscale.
  3. Coding traffic goes through one hub. Explore/plan can pin elsewhere (I pin those to gpt-5.6-luna); the local default is Flash Next.

Default launch is launchd com.ghost128.homebase.flash38, with a rollback script if the new default misbehaves.

Failure modes I hit first

  • Four concurrent streams at 45-65k context dropped decode to ~17 tok/s and looked hung. One GPU was doing the work of a cluster. Caps live in the fan-out post.
  • No MLX DFlash checkpoint for Flash Next. Speculative decode on this model is native MTP. The old DFlash2 path does not apply.
  • Soft product note: this stack powers a lot of the local coding work behind projects I ship, including work featured on destroyrebuild.xyz. Workshop setup.

Series

All ship logs