Short answer
Running inference on a machine you control can keep the prompt and model execution there for a specific configured workflow. It does not prove that every app, plugin, web search, model download, backup, remote access tool, or log stays on that machine. Conversely, cloud services differ in training, retention, account, region, and governance controls; it is inaccurate to say that all cloud providers train on customer prompts. The useful question is: where does this exact file or prompt travel, who can access it, and how long is each copy kept? PebbleRack has not audited a finished hardware/software workflow.
Best for
Local processing is worth exploring when you need offline operation, want to minimize disclosure to external processors, or must place a particular dataset under your own physical and administrative control. This may include personal notes, household records, source code, or business documents, provided you have permission and a secure host. Managed cloud may be better when an organization already has approved enterprise controls, centralized audit, data-loss prevention, contractual terms, or compliance requirements that a home lab cannot meet. A consumer laptop with full-disk encryption and a trusted local app may be a better starting point than a separate server if only one person needs access.
Why use local
A genuinely local-only flow can avoid sending prompts to a model API. For example, LM Studio documents that its downloaded-model chat and local document chat can run offline, while search and model downloads require connectivity. Test by disconnecting from the internet after downloading the runtime and model; a task that still works offers evidence for that narrow path. Keep the model and app version, network setting, and input type in the record. A later web search toggle, browser extension, remote inference option, or shared document integration changes the data flow.
A private home server introduces its own risks. Users need authentication, encryption at rest, backed-up keys, a patch process, and controlled remote access. A local API listening on a network can be queried by another device if access is not constrained. Logs may retain prompts or filenames. Backups may copy sensitive data to an offsite location. A local AI agent with file or shell tools may read or change more than the prompt itself. “On my hardware” identifies the operator; it does not prove sound security.
When not to use local
Do not put regulated or highly sensitive data into a new self-hosted workflow merely because a vendor calls it private. The business may require approved retention, deletion, access logging, incident response, and support. A managed service with appropriate contractual controls may meet those needs better. OpenAI's API documentation, for example, states API data is not used for training by default and describes abuse-monitoring logs and retention controls; the exact endpoint and settings matter. AWS Bedrock describes customer-data protections that differ from a consumer chat account. Verify the relevant product, account tier, endpoint, region, and optional features rather than generalizing from a provider name.
Also avoid local processing if you cannot maintain the host securely. A machine that misses security updates or exposes an unauthenticated API can create a worse privacy outcome than a well-governed external service.
Decision checklist
Create a data-flow row for each stage: source file; local app; model/runtime; any embeddings or search component; tool calls; logs; backups; remote access; and deletion. Record whether it sends content, metadata, or only a model download request across the network. Name each operator and its retention policy. Test one non-sensitive document first and inspect network destinations where feasible. Distinguish provider-published policy from your own observed result. Check whether an optional cloud model is selected at the moment you use the app, not only at installation.
Then ask who can administer the host, whether other household members share it, how a lost disk is handled, and how a user can erase a document from indexes and backups. If you cannot answer these, delay sensitive use.
One practical next step
Draw a simple flow for one proposed document task and mark each boundary: device, home network, internet provider, model provider, and backup target. Run a harmless sample with the network disconnected and record what still works. Keep the sheet private; it may reveal network and account details. Use the other guides to secure remote access and test backup restoration before storing valuable data. No waitlist signup is required to use the guide.
Sources
Official sources checked 2026-09-29: LM Studio offline behavior; LM Studio local, remote and cloud model picker; LM Studio local server exposure; OpenAI API data controls; AWS Bedrock security and data protection.