Running an Apple Silicon Mac as a Einsora host

Status: planned (Phases 1–2, sharing in Phase 4). This describes how the agent will use a Mac (ADR-0013). Verify each step once the agent ships.

What a Mac can run

Workload	How	GPU	Who may use it
Model endpoint	MLX or llama.cpp server (`macos-native`)	Metal	Owner. Invited and community users get it as managed inference only.
Notebook	Jupyter with PyTorch (MPS) and MLX (`macos-native`)	Metal	Owner only
Notebook (strong isolation)	Linux container in its own VM (`macos-container`)	CPU only	Owner, invited, community (opt-in)
VM	macOS guest (`macos-vz`)	Metal (paravirtualized)	Owner, invited, community (opt-in). Max 2 macOS VMs per Mac.
VM	Linux guest (`macos-vz`)	CPU only	Owner, invited, community (opt-in)

The trust matrix in tenancy-and-isolation.md decides who may use what.

Install and join: brew install einsora/tap/einsora. Then either run einsora up (Solo), or einsora agent --join <enrolment token from your Lab's dashboard>. The agent dials out, so no ports need opening.
Power:
- Desktops (Mac mini, Mac Studio): set “Prevent automatic sleeping when the display is off” if you want it available overnight.
- Laptops: the agent shares only while on AC power.
GPU memory: macOS caps how much unified memory the GPU may use. The agent reports what’s usable and keeps headroom for you. Raising the cap is an expert setting (sysctl iogpu.wired_limit_mb); the agent never changes it for you.
Disk: keep enough free space for models (often 5–80 GB each). The agent caches each model once and shares it between workloads.

Default: private.
Invite-only or community sharing is set per machine in the dashboard. You can also set:
- an idle time (for example 15 minutes with no input)
- a schedule
- caps for GPU memory, CPU, and disk
When you come back, guest workloads stop within seconds.
Community users never run native code on your Mac. They get managed inference or a VM, no access to your files or home network, and no internet by default.
You can see guest data on your own machine. Einsora tells guests this before they launch anything.
You pay for the electricity. An M-series Mac uses far less power than a GPU server, but check it anyway (power notes).

Symptom	Likely cause
Endpoint fails to start, “out of memory”	Model too large for the GPU’s share of unified memory. Use a smaller or more quantized model.
Machine shows offline	It’s asleep, or the network changed. The agent reconnects by itself.
Third macOS VM won’t start	The macOS licence allows 2 macOS VMs per Mac