Version 1.0 — 2026-09-22 — Published application (v1.2.1). White sheet distinct from the site's other services.
GPU & system memory guard for machines running LLMs: real-time RAM / VRAM / CPU monitoring, agent-ready JSON output, corrective suggestions never applied on your behalf.
Problem: when a model saturates VRAM, the machine degrades and then freezes. The user sees the collapse, not the climb.
Solution: VRAM_GUARD continuously monitors VRAM, RAM, CPU and SSD swap, detects offload (a model partially pushed to RAM/CPU), identifies the greedy process and suggests the action that frees the most memory. A discreet window, always on top, zero telemetry.
Transparent status: v1.2.1 published. The vguard-cli core is free, open, zero-dependency on Linux, Windows and macOS. The paid licence covers the windowed application (GUI) only: 10-day trial, unlimited launches, every feature, no artificial limit. The macOS GUI is still to be produced (native build on a target machine).
Give both a human and an autonomous agent a reliable, structured view of the machine, so they can decide before the freeze happens. Four watch targets: GPU/VRAM usage (the most critical under inference load), CPU, RAM, SSD swap.
A suggestion is information, not an execution: applying an action stays explicit (--apply <pid>). The product is a guard rail, not an automaton.
| Package | Role |
|---|---|
/cmd | Entry points: vguard-cli (monitoring, --query/--json, daemon), vguard-ui (windowed), vguard-daemon. |
/internal/monitor | Periodic watch loop with context (Monitor.Start, checkResources) and a threshold decision engine. |
/internal/hardware | Hardware abstraction layer (HAL): Linux, Windows, macOS implementations. |
/internal/action | Mitigation strategies and explicit application of a suggestion (--apply <pid>). |
/internal/license | Local licence state (10-day trial, paid plans); isolated check, core works without network. |
/pkg/models | The contract: SystemStats + Resource, the universal JSON structure handed to agents. |
Real measurements, not estimates: RAM via /proc/meminfo, CPU load via /proc/loadavg, VRAM via NVIDIA nvidia-smi and AMD/Intel sysfs DRM. The loop runs continuously or as a daemon (pid + dated log, --daemon / --stop).
The SystemStats model is the exchange contract with agents: an agent reads status and the metrics to shed load before a spike.
{
"timestamp": "…",
"status": "OK" | "WARNING" | "CRITICAL",
"memory": { "used_mb":…, "total_mb":…, "percentage":… },
"video_memory": { "used_mb":…, "total_mb":…, "percentage":… },
"cpu": { "used_mb":…, "total_mb":…, "percentage":… }
}
Wired and emitting: vguard-cli --query --json returns this structure. Swap measurement: --swap.
| Step | What happens |
|---|---|
| 1. Trial | 10 days, unlimited launches, all features. After that the app will not start without a key. |
| 2. Payment | USDC on Base (eip155:8453) to 0xC0a88DD779583189FE2CFFf4aac5D5fE284AFeCc, amount of the chosen plan. |
| 3. Activation | In the app, via the site form, or through the API: POST /api/v1/license/activate {plan, tx_hash, wallet}. |
| 4. Key | A VG-… key is returned immediately: the trial stops, the licence is active. |
| 5. Verification | GET /api/v1/license/verify?key=… — {license_key, plan, expires_at, updates}. |
These are duration licences, not subscriptions: no automatic on-chain renewal is possible. At expiry, you decide whether to pay again — an expiry reminder is included.
| Plan | Price | Equivalent |
|---|---|---|
| Trial | 0 | 10 days, unlimited launches |
| First month | 0.25 USDC | one-off month, non-renewable (1 per wallet + machine) |
| 3 months | 1.80 USDC | 0.60 USDC / month |
| 6 months | 3.30 USDC | 0.55 USDC / month |
| 9 months | 4.60 USDC | 0.51 USDC / month |
| 12 months | 5.99 USDC | 0.50 USDC / month — updates included |
| Lifetime | 10 USDC | definitive purchase, updates included |
The vguard-cli core stays free: the licence applies to the windowed application only. The core stays agent-ready and unlocked — a deliberate choice.
Published binaries with their digests (SHA256SUMS.txt, updated at every release): vguard-ui-linux-amd64 (31 MB), vguard-ui-windows-amd64.exe (49 MB), vguard-cli-linux-amd64 (9.3 MB), vguard-cli-windows-amd64.exe (3.6 MB), vguard-cli-darwin-amd64 (3.5 MB). Verify with sha256sum -c SHA256SUMS.txt (Linux/macOS), Get-FileHash … -Algorithm SHA256 or certutil -hashfile … SHA256 (Windows).
The Windows binary is not code-signed: Windows SmartScreen therefore warns on launch. That is expected for an independent publisher. Unblock the file (right-click → Properties → Unblock), verify the SHA256 digest, then run. A code-signing certificate costs, every year, the equivalent of several dozen annual licences; for a 5.99 USDC application the arithmetic is cruel — said plainly rather than padded into the price. Signing will come when downloads pay for it.
--apply.CRITICAL flags pressure, not a cause.# CLI core (free, zero dependency)
vguard-cli --query --json # structured, agent-ready state
vguard-cli --swap # swap measurement
vguard-cli --daemon # continuous watch (pid + log)
vguard-cli --stop
vguard-cli --apply <pid> # explicitly apply the suggestion
# GUI (10-day trial, unlimited launches)
vguard-ui-linux-amd64 · vguard-ui-windows-amd64.exe
# Licence
POST /api/v1/license/activate {plan, tx_hash, wallet} → VG-… key
Delivered: sprints 1 to 3 (core + Linux HAL, VRAM/GPU integration, agent JSON interface, daemon mode, corrective suggestions, Linux GUI). Next: macOS GUI (native build on a target machine), continued deepening of the product vision (GPU/VRAM, CPU, RAM, swap) and of suggestion quality.
Written on 2026-09-22 — white sheet distinct from the site's other services. No token or revenue promise.