01 / typed input
Not retained
Typed answers are not persisted. Explicit grades—not answer text—schedule local SQLite state.
Polyglot for Codex + Claude Code
Learn a language in the pauses between agent turns.
Instead of staring at terminal churn, collect one useful word or phrase every few completed turns. No extra tab. No streak pressure. Just something small that adds up.
5completed turns
1useful phrase
74language directions
0model calls
Real CLI capture
Five completed turns. One phrase worth keeping.
The shipped Stop hook stays quiet for four turns, then returns one small language-learning moment through the same path packaged for Codex and Claude Code.
Read the demo transcript and summary
A real isolated CLI capture invokes five completed-turn Stop events. The first four return no visible message. On the fifth, Polyglot emits Polyglot starter · hello → hallo. The video intentionally shows only the CLI behavior, while the packaged hook is shared by Codex and Claude Code.
When you want deliberate practice
Recall it. Grade it. See it again when it matters.
The terminal review flow is optional and secondary: it turns ambient exposure into locally scheduled learning.
$ polyglot review --pair en-de --direction reverse Prompt: sehen Your answer: to see Answer: to see Grade [again/hard/good/easy]: good
- Local due state
- Next review: tomorrow
- Progress ledger
- seen 1 · good 1 · due 0
Why “tomorrow”? The CLI shown above displays the answer and accepts the grade. The due-tomorrow line is an explanation of the shipped deterministic new-card good interval—one day—not invented CLI output.
Privacy / token boundary
A narrow learning signal, kept local.
Polyglot isolates practice state from the ambient companion. That boundary is deliberate: a phrase can be useful without becoming evidence about the learner.
02 / ambient
Ungraded, then due-only
A fresh pair can show one labelled starter exposure without creating learner state. After explicit review begins, ambient mode reads due state only and never mutates mastery, grades, or intervals.
03 / isolation
No injection
No model or network call; no learner history, prompt, or transcript is injected into an agent turn.
04 / audit
Small enough to inspect
The largest audited ambient line is 85 characters, about 43 tokens.
The catalog behind the moment
Broad coverage. Conservative learning facts.
The 19,281-entry catalog, interactive cabinet, typed recall, and spaced repetition are there when you want deliberate study. They support the ambient moment; they are not the point of it.
German learning-metadata records. German already existed; this is not marketed as a newly added language.
Conservative exact-lemma article/gender facts.
Exact-text clozes. Automated and not native-speaker reviewed.
New reverse directions: PL→EN 264, UK→EN 265, SV→EN 261, EL→EN 256. Each is mechanically derived from existing shipped source records, with no new translation or transliteration claims. Polyglot does not claim an official Duolingo top 20.
Where it meets you
Inside the coding session. Not in another tab.
Codex Desktop, Codex CLI, and Claude Code share the central Stop-hook experience. Pi and Hermes support extends the same idea to more agent workflows.
Codex Desktop + CLI
One plugin
Packaged skill and fail-soft Stop hook with a compact systemMessage.
Claude Code
Same cadence
The validated plugin declares the same optional Stop adapter.
Pi
Extended
The canonical skill and native end-of-turn extension are included.
Hermes
Extended
The namespaced skill, hook, and command load together.
Make the waiting useful