Connect Ollama or LM Studio to a Phone

Network diagram

The phone is a client. The model stays loaded on a computer, and prompts travel from Android to that computer over a network you choose.

Neither route is phone-offline inference. If the network or computer is unavailable, the mobile client cannot generate through that server. Phos's on-device mode is the separate route for airplane-mode use.

Prepare Ollama

Ollama serves its local API at http://localhost:11434/api by default. Confirm the model on the host before changing network settings. Ollama binds to loopback by default; its FAQ documents the OLLAMA_HOST environment variable for network exposure. Using 0.0.0.0:11434 makes the service reachable on more interfaces, so pair it with a scoped firewall rule.

From the host, test the API and list or invoke the intended model. From another trusted device, test the computer's private address. Do not port-forward 11434 directly from the router. Ollama documents proxy and tunnel examples, but a reachable proxy still needs appropriate authentication, TLS, access restrictions, logging policy, and updates for the intended environment.

Ollama states that prompts for local models are not sent back to ollama.com. Cloud models are a separate route. If strict local-only behavior is required, Ollama documents a cloud-disable configuration; verify current settings and logs on the host.

Prepare LM Studio

Start LM Studio's server in the Developer tab or with lms server start. Confirm a model is loaded and decide whether the client uses LM Studio's native v1 API or an OpenAI-compatible endpoint. Current LM Studio versions document API Tokens, with authentication not necessarily enforced by default. Enable a scoped token for shared or remotely reachable environments.

The --cors option is for browser-origin access and LM Studio warns that enabling it can increase risk. A native Android client normally does not need browser CORS, so do not toggle it as a generic network fix. Check bind address, private route, firewall, endpoint family, model identifier, and token first.

LM Studio documents offline operation for downloaded models and its local server. The host can therefore infer without internet, while the phone still needs the local or private network path to the host.

Configure the phone

Use the computer's private LAN address, private DNS name, or VPN address—not localhost. Enter the scheme and port expected by the profile. Keep separate profiles for Ollama and LM Studio so model names, API variants, and authentication are not mixed. If using a token, store it through app-private, device-protected handling and never include it in screenshots or support logs.

Test with a non-sensitive prompt first. Verify the route label, server identity, and model. Then test cancellation, a multi-turn conversation, server restart, phone network change, and host sleep. A client should surface connection and model errors without silently sending the prompt to another provider.

Troubleshoot by response class

Timeout: check Wi-Fi client isolation, wrong address, host sleep, bind interface, VPN route, or firewall. Connection refused: the host is reachable but nothing listens on that address and port. 401 or 403: network works; token or permission is wrong. 404: base path or API family is wrong. Model not found: endpoint works; load or name the requested model. Malformed response: client/server compatibility or streaming format may differ.

Do not disable the whole firewall. Add one narrow rule, test, and remove or refine it. Do not publish private IPs, tokens, prompts, or full proxy configuration in a public support issue. Technical diagnostics should be sufficient without sensitive content.

Remote access without public exposure

A private mesh VPN is usually the simplest remote extension because it provides authenticated network membership and stable private addressing without opening the raw service to the internet. Restrict the server to the VPN interface or firewall range. If a reverse proxy is required, use TLS, strong authentication, request-size limits, timeouts, rate controls, safe logs, and updates. Avoid weak shared passwords in URLs.

Cloudflare Tunnel and similar tools can make a service reachable, but tunneling is transport, not a complete authorization design. Apply identity-aware access before the inference endpoint and verify that streaming works without exposing prompts in intermediary logs.

Privacy and operating checklist

Keep operating systems, inference server, and model runtime updated. Review host logs and backups. Lock both devices, restrict guest Wi-Fi, rotate tokens after loss, and remove old client profiles. Minimize memory attached to remote prompts. Use phone-local mode for the most sensitive work that a compatible local model can handle.

A personal server is valuable precisely because it offers a middle route: more hardware than the phone, more control than a hosted provider. That benefit survives only when the network and server are treated as part of the privacy boundary.