No account. No cloud. No telemetry. That is the whole product position, and it is easy to claim — so everything on this page is the evidence for it: the model, the runtime, the numbers, and what happens on the wire.
"Advanced AI" is worth nothing as a claim. Here is the whole stack, named, so you can verify it yourself — check the process list, check the firewall, check the folder.
Shipping a gigabyte of model weights inside a system utility would be indefensible. The installer stays small; the model is downloaded once, only if you open Ask, and only after you say yes. Never touch the feature and it is never fetched.
That download is the only outbound connection RIVET makes in its lifetime. After it lands, inference happens in-process on your CPU. Pull the network cable and Ask keeps answering.
Already running Ollama? Point Ask at localhost and it uses whatever you have loaded instead. With nothing listening and no model downloaded, it falls back to a knowledge base compiled into the app — so the feature never dead-ends.
Most tuning utilities fire a command and hope. In RIVET nothing executes directly. Every change — a killed process group, a disabled startup entry, a stopped service, a registry tweak, an emptied folder — is constructed as a reversible action first, and that object is what runs.
Before anything is written, the action records what is there now — the registry value, the service start type, the file location. The undo is built before the change is.
You see the exact set of changes staged before they run. Cleanup reports per-row sizes and waits. Nothing is silent, and nothing runs on a schedule you did not set.
System-level tweaks create a restore point before applying. Deletions route through the Recycle Bin rather than straight out. Elevation is requested per action, not held.
The action runs and writes itself to the log with a timestamp, the target, the prior value and the method that reverses it. The log persists between sessions.
Not just the last one. Because every entry carries its own undo, you can reverse a change from three weeks ago without touching the ones after it.
The point isn't the undo button. It's that a tool you can't break is a tool you'll actually experiment with — which is where the performance you were looking for was hiding.
RIVET started as a single PowerShell script that drew a HUD in a console window. Everything it does today it earned by being rewritten — three times.
Written to stop retyping the same maintenance commands. Console output, no UI, no state.
Still one file, but with a drawn HUD, typed commands, tasks and a rank ladder persisted under C:\rivet\. It sold on itch.io and hit the ceiling of what a script can be.
WPF on .NET 10, one signed MSIX on the Microsoft Store. Eleven modules behind a sidebar, an action log underneath all of them, and an assistant that reads live system state.
Not screen-scraped command output dressed up as a UI. The app talks to Windows through the interfaces Windows provides, which is the only reason the undo can be trusted.
RIVET can read your process tree, your installed software, your startup entries and your disk. Any usage analytics collected from a program with that reach would be worth a great deal to somebody, which is exactly why there aren't any. There is no analytics SDK in the package, no crash reporter phoning home, no unique install ID, and no server on the other end to receive any of it.
No idea which module you use most. Features get built from what people write in, not from a dashboard. That is slower, and it is the trade we chose.
Tasks, log, settings and progression live in your own profile in plain files. Delete the folder and RIVET forgets everything. Nothing is mirrored elsewhere.
Block RIVET at the firewall and use it for a week. Nothing degrades except the one-time model download. The technical notes list every path and key it touches.
The trial is the whole app, and the fastest way to check any of these claims is to watch it in Task Manager yourself.
▸ Get RIVET — free