Mora

Submitted to Anthropic × Cerebral Valley Built with Opus 4.7 hackathon (Apr 21–28, 2026). Demo video and full case study at https://www.youtalk.jp/mora.

An iPad-first, on-device learning app for children with dyslexia learning English as a second language.

Demo video

Three-minute walkthrough of one A-day session on a real iPad: https://youtu.be/zrsgP30Miqg

Status: Early development. Scope, architecture, and APIs are expected to change.

Requirements

• macOS with Xcode 15+ (iOS 17 SDK)
• XcodeGen — brew install xcodegen
• swift-format — brew install swift-format

First-time setup (clone-fresh)

The wav2vec2-phoneme CoreML model is hosted as a GitHub Release asset (see tools/models.manifest), not in Git. After cloning:

bash tools/fetch-models.sh

This downloads + SHA-verifies the model into Packages/MoraMLX/Sources/MoraMLX/Resources/wav2vec2-phoneme.mlmodelc/. Re-runs are no-ops when nothing has changed. The Mora Xcode target also runs tools/fetch-models.sh as a preBuildScript so manifest bumps are picked up automatically.

CI does the same via actions/cache keyed on the manifest hash, so model download bandwidth stays at zero on cache hits.

Getting started

git clone git@github.com:youtalk/mora.git
cd mora
bash tools/fetch-models.sh     # see "First-time setup (clone-fresh)" above
xcodegen generate
open Mora.xcodeproj

Then build the Mora scheme against an iPad simulator.

Repository layout

• Mora/ — iOS app target (thin shell).
• Packages/ — local Swift packages containing the actual logic (MoraCore, MoraEngines, MoraUI, MoraTesting, MoraMLX).
• docs/ — design specs and implementation plans.

License

Mozilla Public License 2.0. OSI-approved, App Store-compatible, file-level weak copyleft.

Note: the yokai asset forge under tools/yokai-forge/ depends on non-commercial upstream models (FLUX.1-dev, Fish Speech S2 Pro). Any generated portrait or voice clip inherits that restriction, so a future commercial release would require regenerating those assets with commercially-cleared models. See tools/yokai-forge/README.md § "Licensing — commercial release requires swap-outs" for the swap-out checklist.

The bundled wav2vec2-phoneme.mlmodelc is derived from facebook/wav2vec2-xlsr-53-espeak-cv-ft (MIT-licensed); attribution preserved in Packages/MoraMLX/Sources/MoraMLX/Resources/.