BackBrain — frequently asked questions

BackBrain is an on-device AI assistant for Windows that actually performs your tasks — handling tax portal work, downloading bank statements, and sending reports on WhatsApp — by operating your own computer, without sending your data to the cloud. An Android companion handles on-device mobile execution.

Chatbots answer questions and write text. BackBrain takes action: it operates real websites, handles logins and OTPs, downloads files, and sends messages on your behalf, all locally on your device.

Yes. BackBrain keeps your credentials in BackBrain's built-in encrypted vault on your own device, which never syncs online. It signs in locally and asks for your approval before any critical action.

No. BackBrain runs on your own computer and processes your data locally. Your passwords, files, and financial details are not uploaded to any external server.

BackBrain is free during its private beta and requires no subscription. It runs on hardware you already own.

BackBrain is a desktop application for Windows with a full Android companion for on-device mobile execution. It also reads one-time passwords from your phone to complete logins.

Yes. BackBrain operates the official tax or government portal on your computer, enters your details, shows you a preview, and submits only after you approve.

Yes. It signs into your bank's online portal locally, reads the OTP from your phone, downloads the statement with proof on disk, and forwards it to whoever needs it after your approval.

No. BackBrain pauses for your explicit approval before any critical or irreversible step, such as submitting a filing or sending a file.

BackBrain is in private beta. Email hello@backbrain.in or reserve your spot at backbrain.in to get early access for Windows and Android.

Yes. Credentials are encrypted on your device with Windows DPAPI — Windows' own encryption — and are supplied just-in-time at the moment of use. Raw secrets (passwords, OTPs, PINs, tokens, card numbers) are never written to logs or transcripts, and screens showing credentials are never sent to a cloud model. If you prefer not to store a password at all, session-only entry holds it in memory for one task and clears it afterward.

Tasks on websites naturally need the internet, but the intelligence doesn't have to: with a local model (for example, Ollama), reasoning runs on your own hardware. If your cloud AI provider has an outage mid-task, BackBrain retries briefly, fails over to any other provider you've configured, and — if everything is unreachable — keeps the run alive rather than dying: it re-probes for recovery, lets you take over the screen while it watches and records, and finishes on proof. Tasks it has done before can replay from local memory with zero AI calls at all.

No. BackBrain never signs you out of a site or app unless you explicitly ask it to. Your accounts stay logged in so the next task starts faster, and a built-in guard blocks stray logout clicks even if the AI considers one. Secure portals like banks still time out on their own schedule, as they always do.

BackBrain pauses and gets the code the fastest safe way available: auto-relayed from your paired Android phone's notifications, read from your linked email, picked up from a Windows notification, or from a code you just copied. You can also type it in directly. OTPs are used once, in the moment, and are never logged or stored.

Yes, three ways. Say 'watch me do this' and demonstrate — it observes without acting, summarizes what it understood, and saves only when you reply 'save' (what you type into fields is never recorded). Say 'remember: …' in chat to give it a durable instruction for a site. Or just correct it mid-run — type a hint or take the mouse — and your correction becomes part of how it does that task in future.

It will not delete files without your explicit instruction plus a confirmation — ever. It will not send, submit, pay, or share anything externally without approval. It will not make a purchase without a pre-authorized mandate (merchant, amount cap, expiry). It will not use an OTP outside an approved session, and it cannot approve its own risky actions — those gates live in a layer the AI cannot touch.

Only from verified evidence. When a step provably works — the page advanced, the file landed on disk — that step is saved to on-device memory. Tasks that succeed repeatedly are compiled into procedures that replay instantly with zero AI calls. Failed runs contribute too: the steps that were verified before the failure fast-forward the next attempt. Learning stays on your machine, and periodic integrity scans clean out bad or misleading memories so mistakes don't compound.

Yes. Every run keeps a full, searchable transcript — every message, every action, every outcome — in the History tab, with secrets redacted. Underneath that, actions are written to a tamper-evident, hash-chained audit ledger, so the record itself can't be quietly altered.

BackBrain doesn't run pre-recorded scripts, so there's nothing brittle to break. It reads the live page on every step and decides the next action fresh — and it clicks controls by what they say on screen, which survives most redesigns. If a remembered procedure no longer matches the page, it hands back to live reasoning mid-flow and re-learns the new path.

Your choice. Bring your own key for Azure OpenAI, OpenAI, Anthropic, Gemini, or Perplexity — or run a fully local model via Ollama, LM Studio, or any compatible endpoint. You can configure fallback chains across providers, and a local-only privacy mode blocks cloud egress entirely. Either way, repeated tasks need less and less model time as BackBrain's on-device memory grows.

Never on its own. File deletion is hard-blocked unless you explicitly instruct it — with words like 'delete' — and then confirm the exact files when prompted. 'Clean up' and 'organize' never imply deletion: organizing means renaming and filing, not removing. This rule is enforced in a layer the AI cannot override.

No. Chatbots produce text; BackBrain produces outcomes. It's an installed operator that works your real browser, desktop apps, files, documents, email, and phone — and reports back with proof. It chats with you along the way, but the chat is the steering wheel, not the product.