Docs · Task 12 · Docs for 1.0.12

Connect a cloud provider beta

Cloud works when a key or a Pro entitlement is present. A build with neither shows the option closed and names who can open it, rather than hiding it. Connecting one is an administrator task, done outside the app — this page is for whoever does it. If you are not that person, the short version is the first section.

⚠
Cloud is beta in 1.0.12, and this is what that word is doing here. The service is deployed and the shipped app can authenticate to it — both measured. What we have not yet done is watch a complete request go out, get served, and come back on a release build. Until we have, treat the cloud path as untested rather than broken: the plumbing is real and the end-to-end run is unwitnessed. This page will say so until that changes.

What you see when it is not connected

The app's Settings screen with the compute routing card. The rail lists
                Profile, Privacy and data, Memory and AI models. The routing card names
                where work runs and shows This device as the current setting, with the
                cloud option listed beside it rather than hidden.
Where the choice is made. The routing card names every place work can run and shows which one is current — here This device, the only row reading Available. The cloud row carries No provider connected. A build with no key listed the option and named who can open it, rather than hiding it. This is the gate shown closed — it is not a picture of cloud serving a request, and nothing on this screen reports whether the service is reachable. Captured on v1.0.12.

The cloud option is present in the routing card and unavailable, carrying the label No provider connected and this explanation:

Nothing in this build is wired to a cloud provider, so a job sent to the cloud would be refused rather than run. There is no switch here to turn on — the option opens when a provider is connected.
i
If no provider is connected, there is nothing for you to enable. The option is shown closed on purpose: an absent option tells you nothing is possible, while a refused one tells you what would make it possible, and who can do it. This one names your administrator.

What an administrator connects

Provider configuration comes from one of two places, and never from the interface: a JSON file in the application's own data directory, or the environment the app is launched in (ARMORE_CLOUD_ENDPOINT, ARMORE_CLOUD_API_KEY, and the optional ARMORE_CLOUD_PROVIDER, ARMORE_CLOUD_REGION, ARMORE_CLOUD_MODEL_ID). There is no in-app form, by design — a provider credential is not something the interface collects.

That second path is open to you today, not only to an administrator in a managed fleet: bring your own endpoint and your own key, and the cloud option opens against the provider you chose and bill. Any OpenAI-compatible endpoint works — a hosted API, or something you run yourself such as vLLM or llama.cpp on a machine with more memory than the one in front of you. This page is written for whoever does that, whether that is your IT department or you on your own machine.

The endpoint has a default; the key does not. Leave the endpoint unset and the app uses Armore's own inference service. The key is the half that is missing from every shipped build, and it is missing on purpose: a key compiled into a desktop binary is extractable by everyone who installs it, and the endpoint it opens bills by the second. So supplying a key alone is enough, supplying an endpoint alone is not, and with neither the reason stays cloud-not-provisioned rather than changing to something that implies progress.

A Pro entitlement opens it with no key at all — and only for Armore's own service. An entitlement is Armore vouching for your device against Armore's proxy; it says nothing about a third party's billing endpoint, so if an administrator points the app at their own endpoint, that one still needs its own key. Two things must both hold: the entitlement file has to be present, and its plan has to verify as paid against the pinned issuing key. Presence alone would be a hole — an empty file would otherwise open a paid option.

⚠
A file being present is not the same as it working, and the app cannot tell you which you have. The service is the only party that verifies an entitlement, so the app reports a document as present, never as valid or active. An entitlement the app accepts and the service rejects is refused at the far end, with the reason coming from there. That is the deliberate choice: hiding the option whenever we cannot verify locally would reproduce the defect this whole page exists to avoid — a silent absence you cannot read.

What choosing cloud actually means for your text

Armore runs on your device by default. No cloud sees your data unless you send it there — and the cloud tier is the case where you send it there, on purpose, one request at a time. This section is the part you should read before you do.

⚠
On the cloud path, your prompt is readable by the people running the service — including us. The request travels over HTTPS, but the encryption ends at the service, not inside a sealed enclave. Your text is then in the memory of a machine Armore operates on Google Cloud Run, in the asia-south2 region. The app grades this tier accordingly and does not claim otherwise: it is the lowest of the sovereignty grades it can report, and a stronger grade would be refused for want of evidence we do not have. If that is not acceptable for a particular piece of text, the answer is to not send that text — which is why the choice is per request and not a setting you flip once.

None of this is true of the device tier or of your own paired devices, and the distinction is the whole point of having two tiers. Work that runs on This device never reaches a server. Choosing cloud is a decision to trade that for a larger model than your machine can hold.

Local-only stops it before anything is sent

With local-only on, a cloud call is refused egress-blocked. The name is precise: the guard refused before any byte reached the wire, not after a failed attempt.

✓
An internal fault is a different refusal — egress-authorize-failed — and the separation is deliberate. If a crash in the policy layer also read as egress-blocked, the product would be telling you it protected you when it had merely broken before deciding anything. A privacy assurance it did not earn is worse than none.

Limits exist, and they are not shown to you

There are two, and they belong to different people. An allowance is yours and resets. A ceiling is ours and clears when we raise it. When one stops a job the card says which, so the product never blames you for our budget — cloud-spend-cap-reached reads "Not your allowance and not your fault."

i
You will not find a figure anywhere in the product, and that is deliberate. No price, no currency, no cap value. The routing card's own rule: a number there "would be the first thing users read and the last thing we can stand behind." This page keeps that rule rather than quietly breaking it.

One honest bound on the ceiling: it is checked on this device, before a job is sent. The metering service counts what each device uses, but it does not refuse on money — so treat the ceiling as a brake the app applies, not as a guarantee enforced somewhere else on your behalf.

The four refusals, and who can clear each

ReasonWhat happenedWho can clear it
cloud-not-provisionedNo provider is connectedYour administrator
cloud-grade-unprovenA provider answered but never proved it met the barThe provider
cloud-allowance-exhaustedYour allowance is spent; it resetsYou, on your renewal date
cloud-spend-cap-reachedOur own ceiling was reachedUs

Three of those four are nothing you can act on, and the page says so rather than offering a remedy that is not yours to apply.

What you can check afterwards

A run that used a provider says so, and one that only partly did says which half ran where — see one run, two halves. If a half came back from a host the product could not verify, the run reads unverifiable rather than local: it will not tell you your data stayed on your device unless it can prove it did.