Lyklaborð

Íslenskt lyklaborð sem skilur íslensku.
A privacy-first iOS keyboard for Icelandic — that also types excellent English.

lyklabord.solberg.is · Download free on the App Store

_{The landing page's hero is a real-time 3D render of an Apple Extended Keyboard II Ð-keycap — tap it, it presses.}

Lyklaborð is what SwiftKey should have been for Iceland: one Icelandic layout that fluently blends Icelandic and English as you type, morphology-aware autocorrect built on BÍN (3 million word forms — the first keyboard on any platform that actually understands Icelandic inflection), and a hard privacy guarantee: the keyboard extension contains zero networking code. Verify it — that's why the source is here. An optional subscription, Lyklaborð+, adds on-device learning with a personal dictionary you fully control.

Base keyboard free forever. Open source. No account for the free tier. No telemetry. No AI bloat. Personal-vocabulary learning is an optional subscription (Lyklaborð+) — see below.

Why

Apple's Icelandic keyboard has no real autocorrect (common words get "corrected" into rare ones). SwiftKey treats iOS as a second-class platform and phones home. Every quirk of both is documented in research/swiftkey-frustrations.md — this keyboard is designed against that list.

What's different

• Icelandic + English on one layout — a two-lane language model absorbs slettur (one-off English words mid-Icelandic) without flipping languages, but follows you decisively when you actually switch. Never hijacks your keyboard mid-sentence.
• Morphology-aware: all 3.70M BÍN word forms are valid vocabulary via a memory-mapped binary (built with lemma-is); the ~110MB model remains file-backed and demand-paged via mmap.
• Under-corrects by design: a word that's valid in either language is never auto-replaced; the literal token you typed always sits in the suggestion bar (quoted) as an escape hatch; URLs, emails, and dotted tokens are never mangled.
• Spacebar near-miss correction: smelirna → smellir á (the spacebar's neighbors are hypotheses, not typos).
• Learning you own (Lyklaborð+): words are learned on-device (2 distinct days, or instantly when you tap the verbatim suggestion), individually deletable — deletions stick — and importable from a SwiftKey data export. Nothing you type in password, URL, or email fields is ever recorded.
• SwiftKey muscle memory: . key with long-press punctuation cluster right of space, spacebar cursor control, double-space period.

Architecture

Five local Swift packages, no third-party dependencies beyond the vendored KeyboardKit 9.9.1 (last fully-MIT release; we maintain our own fork):

Package	Role
LemmaCore	mmap reader for the BÍN binary (validity + morphology)
Lexicon	mmap frequency lexicons (unigrams/bigrams, IS + EN)
TypeEngine	spatial model, corrector, predictor, two-lane language model, typing session; start here for the engine architecture and debugging model
Learning	crash-safe event log + personal model (learned words, tombstones, touch stats)
KeyboardKit	vendored UI/layout framework (MIT)

For the language side of the system—BÍN and morphology, the Miðeind stack, lemma-is, corpora, frontier keyboard research, artifact provenance, and the incremental intelligence roadmap—read ICELANDIC_NLP.md.

The extension reads everything from the app bundle and App Group container and never touches the network; the containing app owns compaction, the dictionary editor, and (soon) encrypted CloudKit sync of your personal data to your own iCloud.

Building

brew install xcodegen
xcodegen generate
xcodebuild -scheme Lyklabord -destination 'generic/platform=iOS Simulator' build

Requires Xcode 26+, iOS 18 deployment target. Tests: swift test in each package under Packages/.

The typing engine runs headlessly on macOS — try it:

cd Packages/TypeEngine
swift run -c release type-repl            # interactive: type, see suggestions + language lane
swift run -c release type-repl run Scenarios/core.scenarios   # 81 behavioral regression scenarios
swift run -c release type-repl bench      # latency percentiles on the real language models

Status

Available on the App Store. Layout, blended autocorrect/prediction, learning + dictionary editor, and SwiftKey import all work. See open items in the table-stakes roadmap.

Data & licenses

Code is MIT. Icelandic language data is derived from BÍN (Beygingarlýsing íslensks nútímamáls), © Árni Magnússon Institute for Icelandic Studies, used under the BÍN conditions — see data/ATTRIBUTION.md. English frequency data from SymSpell (MIT). Evaluation sentences from Wikipedia (CC BY-SA 4.0). Full provenance in data/README.md.