1. Install in one line
  2. It finds free servers
  3. Buyers see them, anonymously
  4. Requests arrive matched
  5. You approve in one click

Silicon Liquid Monitor · free, for data centres and large GPU fleets

Install in one line. Your free servers find their buyers.

A free, read-only agent for your GPU machines. You get a console of your whole fleet: health alerts, free whole servers and free windows ahead. When you choose, those free servers show to buyers as anonymous capacity with full specs, and matching requests arrive for you to approve. Nothing is shared until you say so.

Install

One machine, as a systemd service. It runs as an unprivileged user systemd makes for it, keeps its state in /var/lib/sl-agent, and starts on boot. Run the same line again to upgrade.

curl -fsSL https://siliconliquid.com/monitor/install.sh | sudo bash -s -- --token slc_... --key +3tnwq0sCBnBvMurS8EbPZnn99agn9KQl+5BY8Raguw=

Or check the installer first, then try it with --dry-run, which changes nothing:

V=0.3.0 B=https://siliconliquid.com/agent
curl -fsSL --create-dirs -o v$V/install.sh $B/v$V/install.sh
curl -fsSLO $B/SHA256SUMS
sha256sum -c --ignore-missing SHA256SUMS   # must say OK; stop if it says FAILED
sh v$V/install.sh --dry-run --token slc_... --key +3tnwq0sCBnBvMurS8EbPZnn99agn9KQl+5BY8Raguw=   # checks everything, runs preflight, changes nothing
sudo sh v$V/install.sh --token slc_... --key +3tnwq0sCBnBvMurS8EbPZnn99agn9KQl+5BY8Raguw=

The installer checks the release's signature and every file before anything runs; if a check fails it stops and changes nothing. Preflight prints PASS, FIX or SKIP for each requirement, with the command for each FIX. On macOS use shasum -a 256 -c. Uninstall: curl -fsSL https://siliconliquid.com/monitor/install.sh | sudo bash -s -- --uninstall

Agent 0.3.0, pinned: nothing updates itself. Every release stays at its own address under /agent/v0.3.0/. Release key: +3tnwq0sCBnBvMurS8EbPZnn99agn9KQl+5BY8Raguw=

Put your fleet's token in place of slc_.... No token yet? Sign up below in a minute and these commands come back with it filled in.

Get your token, free

Tell us who you are and we make your fleet: you get its console link and the token your agents use, shown once and emailed. No card, no contract, no charge for monitoring, however many GPUs. We earn only if you sell a free window through us.

Free, with no card and no contract. Your fleet starts private: nothing is shared until you choose to on your console. By creating it you accept the Monitor data terms.

Specs

What it readsGPU model, count and memory; health from NVIDIA dcgm-exporter (DCGM) or nvidia-smi (NVML): XID errors, memory faults (ECC, row remapping), NVLink and PCIe errors, throttling, temperature and power; the interconnect where the machine shows it (InfiniBand node GUIDs and the fastest active link); and whether each whole server is free or busy, from Slurm or Kubernetes (node state and GPU counts only).
What it never readsYour workloads: no process, container, pod, namespace, job or user names. No data, model weights or network traffic. No customer names. Host names go as hashes unless you ask otherwise.
NetworkOutbound HTTPS only (port 443), to siliconliquid.com: one send every 5 minutes. No inbound ports, and it takes no commands. Offline, it keeps up to 72 hours in a local spool and sends it later.
Access it needsNo root, SSH key or BMC access (one systemd step reads the machine's DMI UUID for it). As a service it runs as a user systemd makes for it, with a read-only view of the system. On Kubernetes its permissions are get and list only.
Resource useOne Python file with no dependencies. It reads every 15 seconds and sends every 5 minutes. On Kubernetes it requests 10m CPU and 32 MiB of memory, with a 128 MiB limit.
Runs onLinux with systemd and Python 3.8 or later (the installer needs curl or wget). Docker: an image you build on your machine from the checked release. Kubernetes: with the NVIDIA GPU Operator, by Helm chart (Helm 3) or plain YAML, one agent per GPU node.
GPUs and driversNVIDIA GPUs, through the NVIDIA driver and dcgm-exporter, or nvidia-smi where no exporter runs (then XID errors are not seen). No minimum driver or Kubernetes version is set: preflight checks what you have and prints PASS, FIX or SKIP, with the command for each FIX.
ReleasesAgent 0.3.0, signed. Each release has its own address and never changes; its manifest is signed with our Ed25519 key and lists every file's SHA-256. The installer checks both before anything runs and stops on any mismatch. Nothing updates itself.

The exact payload, data terms and answers for a security review: What Monitor sends, and never sends. --print shows the payload on your machine and sends it nowhere.

Who sees what: you choose

  1. Privatethe default

    Only you see your fleet. We use the data only to run your console and reports.

  2. Anonymous

    Your free servers add to a market count by region and GPU, with no name, shown only where at least 5 fleets add to it and at least a week late.

  3. Listed

    Each free window you confirm shows to buyers with full specs: GPU, servers, dates, region and evidence, never your name. Matching requests come to you to approve.

Change the level, or turn it back to Private, at any time on your console. You can also send one lender or buyer a read-only link that ends on the date you choose.

How free capacity is counted

Never as a utilisation percentage. Idle GPUs scattered inside busy machines cannot be sold, so Monitor counts two things only:

A free whole server

  • all its GPUs healthy and reporting;
  • unallocated in your scheduler: Slurm says IDLE (not MIXED, ALLOCATED, RESERVED, DRAIN, MAINT or PLANNED), or Kubernetes has no GPU requested on it and it is not cordoned;
  • not inside a reserved stretch of your calendar.

With no scheduler to ask, a server counts only if every GPU sat at 5% or less for 24 hours, and it is marked “inferred”.

A free window

  • at least seven days;
  • the same servers free throughout, on one fabric island;
  • from your reservations or a calendar you upload, and confirmed by you.

History from before you installed never counts, and the first contact is our server's clock, not your machine's.

No agent? Two other ways in

A calendar CSV

Upload your free and reserved stretches on the console after you sign up, or paste Slurm's scontrol show reservation. One line per host or island:

host,start,end,status
gpu-01,2026-11-01,2026-12-01,free
gpu-02,2026-11-01,2026-11-15,reserved
island:leaf-a,2026-12-01,2027-01-01,free,8,H100,8

Times are UTC; the end is exclusive. Host names become the hashed ids an agent would send and are not stored. Without an agent, capacity is listed as Declared: what you have told us.

Prometheus remote-write coming soon

Point your Prometheus at Silicon Liquid with a relabel allowlist, and send dcgm-exporter's GPU series with no agent at all. Not open yet: ask, and we will tell you when it is.

The market map

Free whole servers across fleets that chose Anonymous, by region and GPU, as bands. A cell appears only when at least 5 fleets add to it, and the data is at least a week old.

Not enough fleets yet to show any cell without pointing at one of them.

Stuck, or would rather do it on a call? Email arpit@siliconliquid.com: a person answers within one business day. Have spare GPUs but no fleet to monitor? List them on /sell.