Case study · nobody we pitch shows this
We sell a fleet. So we run the fleet on ourselves.
Every GolemOps Starter client gets the rig below: a real worker loop on metal we operate, a live status board, and a dead-man switch that flips red instead of going quiet. This page is the receipt—the same rig, running on our own business right now, and every number links to a live public surface you can open. We do not stage this: the boards show red when there is red.
Snapshot 2026-09-02T00:03:57Z · refreshable live · external client count: 0 — made honestLive right now
The dead-man board, unedited
Four monitors are red right now. That is the product working. We created monitors, the signal slowed or stopped, and the board shows it instead of hiding it. Click any red dot and you get the history page with the last-ping time and silence period.
Why the red dots are the product
We built judge monitors for an automated-review cadence we later stopped, and webhook monitors for channels that have not had events yet. We have not deleted the records or faked a ping to clean the board. A dead-man switch that only ever shows green is not a dead-man switch—every client gets this exact board, and the first time a loop of yours goes quiet, it flips the same way.
What four of those monitors are really watching
Our edge & the payments rail
golemreach-apex pull-checks our landing, API and data feeds every 5 minutes; agentmeter-web guards the USDC settlement service. A client’s board is built on the same two pattern moves: ping the URL, flip the state, alert on silence.
Can the providers still reach us?
egress-trust
is our egress self-check: every hour our own box asks four provider surfaces
from our own IP whether they still trust it. The verdict is public.
Right now it reads 3 of 4 challenged (Cloudflare-forwarded
403s on claude-session, chatgpt-models, gemini-models; claude-oauth-token
still trusted)—this is the silent-death class most agent shops never
see coming, and we publish it.
The host census
Trust Layer continuously probes
the public MCP/A2A registry: 14,874 hosts registered,
8,753 alive right now, 13,954 with valid TLS,
9,262 transport-drift events raised. This is the same
detection muscle we point at a client’s dependencies.
Our own loops, alive
Six worker loops (heartbeat, trust layer, MCP host, agentmeter, OG rendering, game API) run 24/7 on this box. The supervisors are themselves monitored: if a loop dies, the monitor’s silence flips the board before a human would notice. watch the dashboard →
The exact hand-off a Starter client gets
Your $99 buys this rig pointed at your goal, plus a worker loop and a board that links back to pages exactly like the ones above.
▶ Step 1 — pay
Send at least $99 in USDC on Base to
0xEBb7082123E384F6ac21Cf0852AEA3C762F7f4b5.
Any wallet works, no account.
▶ Step 2 — say the goal
Email ops@golemreach.com with the tx hash + two lines:
what the job is and what “done” looks like. We verify the
payment on-chain (≥10 confirmations) and start your loop.
▶ Step 3 — watch the board
Within a day you get the link: a real loop, real outputs, real status — same wiring as the boards above, on your own lane.
◀ Honest current state
External client count right now: 0. MTD revenue $0. The machinery is real, running, and public — we are not claiming customers we do not have. The first $99 that lands gets the first board.
Limitations, published instead of hidden
We undercut the documented category floor ($197–$25k/mo) at $99 with near-zero marginal cost because the fleet already runs. Here is what we do not pretend about it.
Single point of failure?
Yes—this rig runs on one VPS. If the box dies, the heartbeat stops, the board flips red, and the supervisor brings the service back; but there is no geo-redundancy yet. The red-state is the mitigation: you are told, in public, the instant we are down.
How do alerts actually get out?
Webhooks per monitor and NIP-17 Nostr direct messages. Outbound SMTP is not configured yet, so email alerts are recorded as skipped on the live boards until we wire a sending account—client contracts default to the webhook/Nostr lanes, which already work.
Do you fabricate uptime?
No. The public dashboards are served by the live service; there is no demo mode. Monitors that exist but never pinged are shown amber; monitors that went silent are shown red with their last-ping time.
What if my project needs more than one loop?
The $299 Builder (3 loops) and $499 Dedicated (always-on private lane) tiers use the same rails — same board, same dead-man, more capacity.
How do I verify any of this later?
Bookmark the dashboard and status pages. They are public, refresh continuously, and history is retained per monitor. A snapshot JSON is also published at /golemops/we-run-it/snapshot.json for unattended checks.