Cloud model, local model, or both — BackBrain's model flexibility

By Arjun Mehta · Updated 2026-07-05

Operator vs intelligence: a key distinction

BackBrain separates two roles that most AI tools conflate. The operator is the part that drives your screen — clicking, typing, navigating, downloading. The intelligence is the model that reasons about what the operator should do next. BackBrain is the operator; you choose the intelligence.

This matters because it means you can change, upgrade, or replace the model without changing anything about how BackBrain operates your computer. You are not locked into a single vendor's model any more than you are locked into a single vendor's monitor.

Bring your own cloud key

Plug in your own API key for Azure OpenAI, OpenAI, Anthropic, Gemini, or Perplexity. Your key is encrypted in BackBrain's built-in encrypted vault on your device and used only when BackBrain needs to make a reasoning call. You pay the provider at your own rate; BackBrain adds no markup or cloud layer on top.

Because BackBrain compiles successful procedures into on-device memory, many repeat tasks run with zero or minimal AI calls — reducing your model spend over time even if you use a cloud provider.

Run a fully local model

Point BackBrain at a local model served by Ollama, LM Studio, or any compatible endpoint, and all AI reasoning runs on your own hardware. Nothing about the task reaches any external network. For users who work with sensitive financial, medical, or personal information, local-only operation removes the question of what the model provider sees — because it sees nothing.

Performance depends on your hardware, but modern mid-range GPUs handle many practical task-execution workloads comfortably, especially once BackBrain's compiled procedures reduce the number of reasoning calls required.

Fallback chains and outage resilience

Configure more than one provider and BackBrain automatically fails over between them when one is unavailable. Transient errors are retried with short backoff before escalating to the next provider in the chain. If a cloud provider's content filter mistakenly blocks a routine step, BackBrain slims what it sends and retries automatically — rather than stalling your task mid-flow.

If every provider in the chain becomes unreachable, the run doesn't die. BackBrain keeps the task alive, re-probes for recovery, and offers you the option to take over the screen manually — all while watching, recording what you do, and learning from your takeover for next time.

Local-only mode: blocking cloud egress

Enable local-only mode and BackBrain blocks all cloud model calls outright. There is no silent fallback to a cloud provider if the local model is slow or unavailable — the run pauses rather than sending data externally without your knowledge. The block is enforced at the network layer, not just a configuration flag the software might ignore.

Redaction before the cloud

If you use a cloud model, BackBrain does not send raw screen content to it. Sensitive data — credentials, account numbers, personal identifiers — is redacted from what is sent. Screens showing credential entry are never transmitted. You can inspect what BackBrain considers before sending in the model-call log.

Which option is right for you?

• Cloud key: best for most users who want capable reasoning, low setup cost, and don't mind a small cloud dependency. Redaction keeps sensitive content local.
• Local model: best for users who work with highly sensitive information and want zero cloud dependency, and have hardware that can support a capable model.
• Fallback chain: best for users who want reliable task execution even during provider outages and don't need strict local-only operation.

For a broader look at how on-device AI compares to cloud-based assistants across privacy, capability, and independence, see our full on-device AI vs cloud AI comparison.