Frequently asked questions

The binary. CrashCart is a catalog of prebuilt, self-contained executables — you pick a provider, model and platform (Windows or Linux), download one file, copy it to a server, and run it. The subscription is what keeps your download current as models and builds move on.

No. Nothing about it needs a broken server. The snapshot it takes at startup — OS, kernel, disk, memory, failed units — is just context, so you skip the part where you paste all that in. On a healthy box it is simply an AI session that already knows which machine it is sitting on. The crisis is what it is built for, not what it requires.

Yes. CrashCart is BYOK — bring your own key. You set CRASHCART_API_KEY in the environment at runtime and the binary uses it for that session. We never see it, it is never baked into the binary, and it never leaves the machine you run it on.

No. You pay your model provider directly for tokens, on your own account, at their rates. CrashCart charges for access to the binaries — the two bills are separate, and you keep full control of the provider relationship and spend.

The provider and model are compiled in, which is what lets the binary be a single dependency-free file with nothing to configure in a crisis. Switching models means downloading a different binary rather than editing config. That is also why plans are tiered by how many models you can download rather than by usage.

A subscription. It gives you ongoing access to the catalog: unlimited re-downloads of the latest build, covering Windows and both Linux architectures, with the tiers set by how many models you can download — one, up to five, or every model — and the higher tiers adding Active Mode and MSP team features. You still bring your own key and pay the provider separately. Current prices, monthly or annual, are in the pricing section on the home page.

Anything you would ask a knowledgeable colleague sitting at that machine. Reading a log you did not write. Working out what an unfamiliar unit or config file is doing. A sanity check before a change, or a look around after one. Getting a new hire oriented on a real box instead of a wiki page. It is a troubleshooting tool because that is when you need it most, not because it refuses to be useful the rest of the time.

That is rather the point. It is one static file with no runtime, no dependencies and no installer, and it needs no elevated privileges — on Linux it reads /proc, df and systemctl as an ordinary user (no root); on Windows it is the same story with no admin rights. Copy it over and run it (chmod +x on Linux, or the .exe on Windows); uninstalling is deleting the file. That makes it workable on locked-down hosts, someone else's server, a short-lived container, or a box where you would never get a package approved.

No. Today it only reads. Chat mode takes the startup snapshot and talks; Active Mode adds read-only tools — system info, disk usage, tail a log, read a file. Nothing runs commands and nothing writes. Command execution is planned, and when it lands it will ask before every command and default to no.

Chat mode talks; Active Mode looks. It lets the model reach for read-only tools — system info, disk usage, tailing a log, reading a file — instead of asking you to run things and paste the output back. It is compiled into higher-tier builds only, so a chat-mode binary cannot use tools even if asked to.

Linux (x64 and ARM64) and Windows (x64) today. It is a static binary with no runtime, no installer and no dependencies — on Linux, copy it over, chmod +x, and run; on Windows it is a single .exe you run directly. macOS builds are planned; nothing about the model needs a native machine to build them — they all cross-compile from one host.

The catalog rebuilds on a regular schedule, so new models and refreshed builds appear automatically — nothing for you to do. How many models you can pull depends on your plan: Solo covers one, Multi up to five, and Pro and MSP every model — always at the latest build.

The binary talks to your model provider with your key, and to nothing else. Prompts and output go where you would expect — to the provider you chose. The portal only records download metadata so it knows what you are entitled to.

Yes — that is what the MSP tier is for: unlimited clients, per-client prompt files, client-tagged download reporting, a white-label binary name, and a bulk download API. Subscribe to it directly from the pricing page — no sales call needed. For volume licensing or anything bespoke, use the contact form.