Ai Owned Host

Use an Ollama-compatible owned host

The recorded Private AI path sends a prompt and selected document contents to an Ollama-compatible host the user configures. It is private-network inference, not on-device inference and not a guarantee that no data leaves the phone. Privacy depends on control of that host, its logging policy, network configuration and model behavior. The app's public-endpoint refusal reduces accidental cloud fallback; it does not prove ownership of an IP address.

Prepare the host

Run an appropriate model server on hardware you control and confirm its local model list using the server's own tools. Choose a model whose license and resource requirements fit the task. The Android client discovers available models from the configured host. The website's model gallery is illustrative, not a tested promise that every listed model works with the recorded candidate.

For a locally connected test device or emulator, the documented path uses adb reverse to reach the host's loopback port. Keep the model server bound appropriately and do not expose it broadly just to make a demo easier. A LAN path should use the required HTTPS configuration and an address allowed by the app's endpoint policy. A public tunnel is not an owned-private-host equivalent.

adb reverse tcp:11434 tcp:11434

Then configure http://127.0.0.1:11434 in the app for that reversed loopback path. This command does not install Ollama, download a model or authenticate a host. It only forwards the Android-side port to the connected machine. If the host is unavailable, the app should show that result rather than synthesize a reply.

Review data before sending

Selected document contents leave the phone for the configured server. The receipt retains digests rather than the document body, but the server receives the body used for inference. Do not upload sensitive material unless you understand the server's retention and access controls. A session-only phone history does not erase logs, caches or backups on another machine.

The client applies bounded reads and strict UTF-8 handling to attached text. Those checks protect memory and decoding behavior; they do not sanitize every semantic instruction contained in a document. If an AI answer later influences a wallet action, preserve a human review boundary. Model output must not become automatic transaction approval or arbitrary node execution.

Understand cancel

Cancel closes the client socket and stops the app waiting for more output. It does not guarantee the server stops generating. A host may continue work after losing a client connection. State this distinction in UI and documentation. A cancelled receipt should describe the client observation without asserting that remote CPU work ceased at the same instant.

Qualify the path

Record model identity, host environment, endpoint, a non-sensitive prompt, response, timings and any errors. Demonstrate public-endpoint refusal separately from successful private-host inference. A refusal is a useful policy test, but it does not prove a model ran. The current public ledger leaves live inference unqualified. Once a real model answers, promote only that bounded configuration and keep on-device inference, arbitrary models and production privacy claims separate.

---

Wiki home · Qualification updates · Sources and credits