// the take
Mural is a native iPhone (SwiftUI) and Android (Jetpack Compose) app that teaches a language through live voice conversation with an AI orb, talking directly to OpenAI with your own API key instead of a hosted backend. It's aimed at developers comfortable building from source in Xcode and paying their own OpenAI bill, not at someone expecting to find it on the App Store.
Learning data genuinely stays on-device — conversations and vocabulary live in SwiftData/local storage with no Mural cloud sync or analytics SDK, which is an actual architectural choice rather than a privacy-policy claim. The BYOK design means the core app needs no inference proxy or server Mural has to run and pay for. There's real CI discipline for a pre-launch project: 70 core tests plus 20 native UI tests on iOS, 79 backend tests, and gitleaks secret scanning wired into GitHub Actions. Each of the eleven language modules carries its own teaching policy and themes, and the recall-strength bars are derived from actual conversation evidence rather than a generic spaced-repetition timer.
There's no App Store or TestFlight listing — the only install path today is building from source with a Mac and Xcode 26, and free-tier (non-paid Apple Developer) installs expire every seven days and need a manual rebuild to keep working. You need an OpenAI project with billing enabled and access to specific models (GPT-Live-1, GPT-5.6 Luna); the app's usage display is explicitly called an estimate and 'not a billing cap,' so a chatty learner can get an unpleasant OpenAI invoice with no in-app guardrail. The hosted/commercial layer (services/api) is half-built and disabled by design — identity, minute purchases, and billing allowances exist in code but aren't live, so the product's actual business model and API surface are still unsettled. Speech quality claims are honestly caveated but real: accent targeting, pronunciation correction, and recognition quality across the eleven languages are explicitly unverified by native/fluent speakers.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →