One agent. Every GPU.
Peak revenue.

Ermeon Agent runs on your servers and connects them to Vast.ai, Clore.ai and mining. One panel to monitor the fleet, manage every listing and configure the hardware — with a strategy engine that routes each GPU to its best net $/GPU·h. Your AI assistant can run it too, over MCP.

$ curl -sL get.ermeon.com | bash

83% fleet utilization · $0.142/GPU·h effective · 16 servers, one panel

Numbers from the live demo fleet — open it below.

Live demo

Don't take the screenshot's word for it

The panel above is running right now at panel.ermeon.com — a sample fleet of 16 servers with per-GPU telemetry, rentals, mining jobs and a Decision Log you can read entry by entry. Sign in and check the math yourself — no install, no signup, nothing to configure.

login
password
Open live demo

Demo fleet with sample data — seeded telemetry and history, no real hardware behind it. Ermeon Agent is in early access; your own servers connect with one command once you're in.

The problem

Your GPUs already earn on three platforms.
You manage them in three tabs.

Three agents, three dashboards

Every marketplace wants its own daemon, its own account, its own rules. Nobody shows you the whole farm.

Money left idle

GPUs sit unlisted while spot prices spike on the next platform over. Idle hours are pure loss — minus the power bill.

Manual everything

Pricing, switching to mining and back, watching for CUDA errors at 3 AM. Your farm runs 24/7 — you shouldn't have to.

Ermeon Agent

First you see everything. Then you control everything. Then it runs itself.

One agent on each server, one panel over the fleet. Five capabilities, in the order you'll actually use them.

01 · Monitor

See every GPU the way the silicon sees itself

Ermeon Agent reads the fleet at the hardware level and streams it into one panel — no more guessing which tab holds the truth.

  • Per-GPU telemetry — core, hotspot and VRAM temperature (GDDR6X memory read at register level), clocks, power draw, and what each card is doing: rented, mining or idle.
  • Rental awareness — live status of every rental on every platform: who's on the machine, at what price, since when.
  • Full event log — rentals starting and ending, miner preemptions, reprices, self-tests, maintenance. Everything that happened, with a timestamp and a reason.
  • Alerts that matter — Telegram: server offline, rental started or ended, CUDA error. Silence when everything is fine.

hotspot 78°C · vram 84°C · 312 W — nothing to report

02 · Manage

Every marketplace, one panel — moves included

Listings, prices and platform assignments for the whole fleet from one screen. When another platform pays more, moving a server there is a single action, not an evening of SSH.

  • One-action moves — Vast ↔ Clore ↔ mining: unlist, wait out active rentals, switch platform agents, relist. The agent sequences it and logs each step; you click once.
  • Zero-downtime fallback — idle GPUs mine automatically; any incoming rental preempts the miner in seconds. Mining profit becomes your price floor, not a separate job.
  • Queued around rentals — disruptive actions like reboots and driver updates wait for the rental to end by default. Your reliability score and ratings never take the hit.
  • One server, one active listing — a machine is listed on exactly one marketplace at a time, so it's never promised to two renters at once.

move Server-04: vast → clore · queued until rental ends 16:40

03 · Configure

Set the policy once. Apply it to one server or the whole fleet.

Hardware tuning and pricing rules live in one place — defined as policies, not per-machine tweaks you'll forget about. The agent enforces them locally even if the connection to us drops.

  • Fans by target temperature — a closed loop on the agent holds your target for core, hotspot or VRAM. During a rental, fan control is never handed to the renter.
  • Power limits with live math — see profit, electricity cost and performance impact recompute before you apply. Applied when idle, locked during rentals. Electricity is the one cost you control: −40 W per GPU at $0.10/kWh is $0.004/GPU·h — small, until it's the whole fleet.
  • Flight sheets — define what idle GPUs mine: image, algorithm, coin. Per server, per group, or fleet-wide.
  • Pricing policies — Market-follow (anchor to live percentiles, your offset), Template (a price profile per GPU class), or Manual (you set it, the engine only alerts).

fan policy: target vram 86°C · applied to group "3090-rigs" (4 servers)

04 · Maximize — Strategy Engine

A strategy engine that shows its work

One metric rules everything: net $/GPU·h after fees. The engine compares expected revenue across rentals, spot and mining — and moves capacity only when the math clears fees and switching cost. Net $/GPU·h isn't a dashboard stat here; it is literally the objective function.

E[rev] = price × P(rent | price, market) × duration
− electricity − fees − switching_cost
  • Anchored to live percentiles — Vast market data refreshes every 5 minutes, Clore every 10–15; listings recompute every 2 hours. Medians and percentiles, never averages.
  • Mining as your floor — the minimum price is mining profit +10%. A GPU never rents for less than it would earn just mining.
  • Server-side guardrails — price moves capped at ±15% per hour, spot never cut more than once per 10 minutes, repricing freezes if market data goes stale. Guards live on the server, not in the UI.
  • Every decision explained — the Decision Log records each action with the inputs it was based on and a rollback plan. A move that doesn't rent out in 12 hours reverts automatically.
14:20Server-11 → mining — Clore spot for 3090s fell below floor: $0.049/GPU·h vs $0.058 mining. Relisting when spot clears $0.064.
09:05Server-03 repriced $0.34 → $0.31 — Vast median for 4090 down 9% in 4 h; holding position at P40. Within ±15%/h guard.
02:11Server-07 driver update queued — rental active until 07:30; running it now would cost reliability. Scheduled for 07:35.
Maximum ProfitMaximum UtilizationMining FirstLong Rentals OnlyCustom

05 · MCP inside

The first GPU fleet your AI can run

Ermeon Agent ships an MCP server. Connect Claude or any MCP-capable assistant and it can watch your farm, analyze markets and — with your permission — operate it. The guards are server-side, not prompt-side: the same price and unlist limits that bind the UI bind the agent. Every call is audited, and money is out of reach by architecture — withdrawal and payout tools simply don't exist.

Claude, why is Server-11 mining instead of renting?
Clore spot for 3090s dropped below your floor at 14:20. Expected rental revenue is $0.049/GPU·h vs mining at $0.058. I'll relist automatically when spot recovers — want a notification when it does?
1 · Monitor

Read-only: telemetry, earnings, Decision Log

2 · Advisor

Proposes actions into your approval queue

3 · Operator

Acts within hard limits: prices ±15%, power and fan policies, mining pause/resume

4 · Autopilot

Everything except money and the irreversible — ever

Tokens are scoped per level, per server list, with expiry. Every act-tool supports dry-run: full effect calculation, nothing applied. A spike in agent activity auto-downgrades the token to Advisor and alerts you.

How it works

From bare Ubuntu to an earning fleet

Install the agent

One command on Ubuntu. Your servers appear in the panel in about a minute — hardware, GPUs, existing marketplace agents and miners auto-detected.

Connect marketplaces

Paste an API key. Preflight checks run, listings import with market-smart defaults. Payouts stay 100% direct: marketplace → you.

Pick a strategy

Maximum Profit, Maximum Utilization, Mining First — or your own rules. The engine does the rest and explains every move.

Pricing

Free forever. We grow when you grow.

Free
$0
  • Up to 4 servers
  • All strategies & optimizer
  • Marketplace analytics
  • MCP: Monitor & Advisor
  • Telegram alerts
Pro
$2 /GPU·mo
  • Unlimited servers
  • Teams & roles
  • ML-driven pricing
  • MCP: Operator & Autopilot
  • Priority support
Enterprise
Custom
  • Datacenters & bare metal
  • SLA & dedicated adapters
  • MIG / vGPU
  • On-prem control plane

$2/GPU·mo over 720 h is $0.0028/GPU·h. If the engine can't beat that, don't pay for it.

FAQ

Fair questions

Yes. Clore explicitly allows background jobs on idle servers; Vast ships defjob for exactly this purpose. Ermeon Agent keeps exactly one active listing per server and queues disruptive actions until rentals end — every platform's rental-time exclusivity is honored, and your machine is never promised twice.
Never. Marketplaces pay you directly — Wise, PayPal, crypto, whatever you've set up. We never ask for wallet keys, and money-related actions don't exist in our API or MCP by design.
Yes — mining stays your fallback income and your price floor. Ermeon Agent decides when mining or renting pays more and switches without downtime; any rental preempts the miner in seconds.
The local agent keeps fan, power-limit and watchdog policies running, and marketplace agents stay up. You lose optimization for a while — not uptime, not rentals, not reliability score.
Ermeon Agent is in early access: we onboard fleets in batches while marketplace adapters harden against real workloads. Monitoring and the panel ship first; automation rolls out to early-access users in stages. The live demo shows the exact panel you'd get — with sample data instead of your servers. Early access is free.
Vast.ai and Clore.ai at launch, plus built-in mining. RunPod and io.net are next — you can vote for integrations right in the panel.

Ready to put every GPU to work?

Ermeon Agent is in early access. Join the list — the first 500 farms get Pro free for 6 months.

Or look around first: Open live demo · demo@ermeon.com / demo-58aa4a3d

✓ You're on the list — we'll be in touch