# Infrastructure inventory (measured 2026-11-09)
Source: live mld scan from the cluster gateway (M1M32) and direct inspection of M5M36 over SSH, on 2026-11-09 ~21:40 ET.
Historical inventories were not trusted. Private addresses, serial numbers and credentials are deliberately omitted.
# Compute nodes (29 Macs on one LAN, 27 reachable at scan time)
| Node |
Model |
Cores |
RAM |
Free RAM |
Load (1 min) |
Free disk |
PM2 apps |
Notes |
| M3U96a |
Mac Studio |
32 |
96 GB |
57.5 GB |
11.4 |
226 GB |
17 |
busiest app node (FinData lake) |
| M3U96b |
Mac Studio |
32 |
96 GB |
79.9 GB |
1.6 |
87 GB |
10 |
low free disk |
| M2U64 |
Mac Studio |
24 |
64 GB |
44.4 GB |
22.9 |
241 GB |
11 |
no GUI session, heavily loaded |
| M4M64a |
Mac Studio |
16 |
64 GB |
29.8 GB |
1.5 |
276 GB |
10 |
hosts agentd (cluster telemetry API) |
| M4M64b |
Mac Studio |
16 |
64 GB |
51.3 GB |
1.2 |
13 GB |
11 |
disk almost full |
| M4M64c |
Mac Studio |
16 |
64 GB |
49.5 GB |
1.6 |
270 GB |
5 |
chat/note apps + Postgres |
| M5M36 |
Mac Studio (M5 Max) |
18 |
36 GB |
24.1 GB |
1.2 |
300 GB |
0 |
selected host — see node-selection.md |
| M4BP48 |
MacBook Pro |
16 |
48 GB |
38.4 GB |
0.8 |
805 GB |
0 |
laptop (may leave the LAN) |
| M4BP36 |
MacBook Pro |
14 |
36 GB |
— |
— |
— |
— |
offline at scan time |
| M4M36 |
Mac Studio |
14 |
36 GB |
23.9 GB |
2.9 |
412 GB |
0 |
reserved for ka guardians |
| M2M32 / M2M32b / M2M32c |
Mac Studio |
12 |
32 GB |
~21 GB |
~1.2 |
257–409 GB |
1–3 |
app nodes |
| m4mc |
Mac mini |
12 |
24 GB |
13.8 GB |
1.5 |
424 GB |
0 |
no Homebrew runtimes |
| M1M32 |
Mac Studio |
10 |
32 GB |
19.3 GB |
2.3 |
83 GB |
2 |
cluster gateway (git server + mld); hosts no apps by policy |
| m4mm / m4mn |
Mac mini (M4 Pro) |
12 |
48 GB |
~37 GB |
1.4 |
~340 GB |
1 |
app nodes |
| m6ma / m6mb |
Mac mini (M6) |
12 |
24 GB |
~15 GB |
1.1 |
~392 GB |
0–1 |
|
| m5ma |
Mac mini (M5 Pro) |
15 |
24 GB |
13.4 GB |
1.1 |
374 GB |
0 |
|
| m4ma / m4mb / m4md / m4ml / m2m16 |
Mac mini |
10 |
16–24 GB |
7–17 GB |
~1.5 |
178–426 GB |
0 |
m4md offline at scan time |
| M3BA24 / M3BA16 |
MacBook Air |
8 |
16–24 GB |
8–14 GB |
1.3 |
~415 GB |
0 |
laptops |
| m2m8a / m2m8b |
Mac mini |
8 |
8 GB |
~3 GB |
1.6 |
195–423 GB |
0 |
too small |
Totals (inventory): 29 Macs, 401 CPU cores, 1 068 GB unified memory. The LAN is 1 GbE (measured in a previous project);
there is no 10 GbE fabric and no NAS in the current cluster, despite the historical hints.
# External server
| Name |
Role |
| BHS64 (OVH, Beauharnois) |
Only external machine. Public gateway of the "MacLustr Tunnel": Caddy (automatic TLS) + WireGuard hub. Public sites are Caddy routes that reverse-proxy over WireGuard to a node port. |
# Mobile nodes
Three iPhones can pull small jobs through maclustr-mobile; they are not SSH nodes and are out of scope for execution.
# Selected host — M5M36 (inspected directly)
| Item |
Value |
| Hardware |
Mac Studio, Apple M5 Max, 18 CPU cores (6 super + 12 performance), 32 GPU cores, 36 GB unified memory |
| Storage |
460 GB volume, 301 GB free |
| OS |
macOS 27.0 (26A425), uptime 20 days at inspection |
| Power policy |
never sleeps, auto-restart after power loss, auto-login (cluster harden applied) |
| Tunnel |
already a WireGuard peer of BHS64 (wg1) |
| Present before this project |
Homebrew, Node 26.10, pnpm 12.6, npm, git 2.54, Python 3.14, uv, PM2 7 (no processes), jq |
| Installed for this project (2026-11-09) |
tmux 3.8, ripgrep 15.2, PostgreSQL 17.11 (brew service), Claude Code CLI 2.1.322 (npm global) |
| Not installed |
gh, pandoc, TeX (install when a research workflow needs them) |
| Claude Code state |
freshly installed, not authenticated (no ~/.claude); authentication is done by the owner inside the first web terminal session |
| Listening services before deployment |
only Apple Remote Desktop agent |
# Cluster services relevant to Claude Remote
| Service |
Where |
Use |
| agentd 3.1.2 |
M4M64a (LAN, bearer token) |
real-time node inventory + metrics (/api/cluster) — the Cluster page's data source |
| mld (maclustr-dispatch) |
M1M32 |
deployment, placement, tunnel routes, self-healing every 5 min |
| spbgit |
M1M32 (gitsrv SSH alias on every node) |
git remotes for all projects |
| skills registry |
skills.spboucher.ai (M3U96b) |
read-only search adapter |
| Local LLM endpoints |
none |
the former llm.maclustr.io gateway (Qwen, DeepSeek, Mistral via MLX) was dismantled on 2026-06-09; no local model is currently served. The Local Models page therefore starts empty. |