Langdogu

This site in

Get Lingua Layer

Where every piece stands

Langdogu is built one piece at a time, and every piece carries one exact label. A label changes only when the piece does. No dates are promised.

How to read the marks

Each label has a dot glyph of its own shape, so a status never depends on colour alone.

Live
Shipped and public.
In development
Being built now, not yet released.
Being rebuilt
Existed in an earlier prototype, now being rebuilt for Langdogu.
Planned
Intended to be built. Nothing of it is released.
Intent
What Langdogu means to do. Not a promise.
Past prototype
Existed earlier and is retired. History only.
Every piece and its statusAs of 10 October 2026

So far

Everything that has happened, in order. The bright dots are the days something changed; the mint one is the day something went live.

  1. Early 2026An earlier ontology prototype, built around an 11-language graph. Now retired.
  2. Mid-2026The method redesigned: a history that is only added to, and languages as data.
  3. August 2026A list-based known-word highlighter, prototyped as a Firefox extension.
  4. Langdogu named; langdogu.com.
  5. Lingua Layer 0.2.1 public on Firefox Add-ons.
  6. Langdogu app development started; the companion designed.
  7. Now

Every piece in detail

What each piece is, what exists of it today, and what done would mean. A piece is done when it works inside Langdogu and its label moves to Live.

Live

Shipped and public.

  • Lingua Layer

    A translation extension for Firefox. It reads Chinese and Japanese into Korean on whole pages, selected sentences, images and manga.

    Today
    Version 0.2.1 has been public on Firefox Add-ons since 9 October 2026, for Firefox 140 or later on desktop and Firefox for Android 142 or later. It needs your own companion service and a Codex CLI signed in with your ChatGPT account.

    Everything Lingua Layer does

In development

Being built now, not yet released.

  • Langdogu app

    The learning app that pairs with Lingua Layer. Lingua Layer handles the moment you meet a word on the web; the app keeps the record and the review.

    Today
    Development started on 10 October 2026, for iOS, Android, desktop (Windows) and the web. It is made first for Korean speakers reading Japanese on the web; Simplified Chinese and English are also in scope. No release dates.
    When it is done
    You could save an expression with its sentence while reading, understand its meaning, structure and examples, and review it with spaced repetition (FSRS). On your next read, the expressions you know would be told apart. No streaks. A local mode without an account is planned.

    More about the app

  • MCP server for AI assistants

    A remote MCP server, so Claude, ChatGPT and other assistants that support MCP can open your Langdogu vocabulary and talk with you about what you learned today.

    Today
    Being built, not released. Connections from claude.ai and chatgpt.com are not yet verified. The tools, in plain words: search your saved expressions; open one expression with its meaning, context sentence and source; a summary for today or this week; the reviews that are due; save an expression.
    When it is done
    You would add Langdogu to your assistant and approve it on the Langdogu sign-in screen, with two permissions granted separately: read (expressions, due reviews, the summary) and write (saving an expression, nothing else). Assistants would get these tools, never raw database access. The same server would run in the managed cloud and in a self-hosted install.

    Open by design

  • Self-hosting

    Run Langdogu on your own PC, NAS or server, with the same learning features, API and MCP server. Supported, and free forever.

    Today
    Being built, not released. There is no download, installer or guide yet.
    When it is done
    You would bring your own AI key and pay the provider directly, with no Langdogu fee. Your own budget would take the place of managed billing.

    Open by design

  • Mascot companion

    An original dot-matrix mascot that you befriend. Its look can change.

    Today
    Designed on 10 October 2026. A demo that lives only in your browser is on the home page; the real companion lives in the app.
    When it is done
    Bond levels would never go down and would be earned mainly by real learning. It would remember milestones, and accessories would be unlocked by learning, never bought. No guilt and no streak pressure.

    Try the demo

Being rebuilt

Existed in an earlier prototype, now being rebuilt for Langdogu.

  • Ontology core

    Concepts above any single language, fixed expressions kept whole, spaced repetition (FSRS) for every item, where each word appeared, and a history that is only ever added to.

    Today
    Designed, and partly prototyped, in an earlier system built around an 11-language graph. Now being rebuilt for Langdogu.
    When it is done
    The parts would work together inside Langdogu as one record of what you know, built up over years from everything you read, watch and hear.

    How the personal ontology works

  • Known-word highlighting

    Sentence-level AI judgment of which words you already know, with level filters such as JLPT, so you can read at the edge of your ability.

    Today
    A list-based version, exact-match highlighting seeded with 2,211 JLPT N5 to N1 kanji, was prototyped as a Firefox extension in August 2026. Sentence-level AI judgment is new. In the first app, “known” is what you mark yourself.
    When it is done
    The words you know would be judged sentence by sentence. JLPT has published no official vocabulary list since 2010, so level filters would use clearly labelled third-party lists.

  • Model routing

    Light models such as Claude Haiku handle preprocessing and most sentences. Idioms and ambiguous passages go to stronger models such as Claude Opus.

    Today
    The earlier prototype routed by cost, with a stronger model spot-checking a 10% sample of a cheaper model’s output. Escalation by difficulty is new.
    When it is done
    Inside Langdogu, light models would handle most sentences, and the passages that need it would go to stronger models, spot-checked.

Planned

Intended to be built. Nothing of it is released.

  • Learning loop

    From reading to review and back.

    Today
    Not in Lingua Layer 0.2.1. The reader on the home page previews saving a word with its sentence.
    When it is done
    You could save an unknown word with its sentence while reading, review it in Langdogu, and the words you know would be marked the next time you read.

  • Video and audio

    Subtitle-based video learning cut by utterance rather than by scene, and audio.

    When it is done
    Subtitled video and audio would share one vocabulary and one review state with everything you read.

  • Pricing

    Planned: local and self-hosted use free forever; Cloud Free for sync and backup; Plus at $8 a month; managed AI at the provider’s cost × 1.3, shown per request, with a default monthly cap of $3 and no automatic top-up.

    Today
    Nothing is on sale. The first beta runs without payments.
    When it is done
    The $3 is a spending cap, not a monthly fee. Cancelling or hitting a limit would never block reading, reviewing or exporting what you already have. Exact Cloud Free and Plus limits would be published before any paid plan starts. Prices in US dollars before tax.

    See the planned pricing

Intent

What Langdogu means to do. Not a promise.

  • Open by design

    Your record of what you know should outlast any one app, including this one.

    Today
    Lingua Layer’s code is All Rights Reserved. It is not open source today. The MCP server and self-hosting are in development (above).
    When it is done
    The core would be open source, with Anki import and export. The managed cloud subscription and usage-based AI would pay for it.

    Open by design

Past prototype

Existed earlier and is retired. History only.

  • Earlier ontology prototype

    An earlier system built around an 11-language graph. The ontology was first designed and partly prototyped there, with rules shared between languages and model routing by cost.

    Today
    Retired. What it taught is being rebuilt for Langdogu as the ontology core, known-word highlighting and model routing.

What gets built next

The order of work, in plain words. It is an order, not a schedule.

The test everything waits for

In development
  1. On the webSave a word with its sentence
  2. On your phoneReview it
  3. On the desktopYour next read shows you know it

The MCP server and self-hosting have to serve this same loop, not a different one.

First

  1. 1

    The core loop, on every device

    Save on the web, review on your phone, see it on your next read. Nothing else is worth connecting until this holds.

    In development
  2. 2

    The MCP server, reading first

    Assistants can search your saved expressions, open one, and see today’s summary and the reviews that are due. Saving an expression needs its own permission.

    In development
  3. 3

    Self-hosting, end to end

    The same learning features, API and MCP server on your own PC, NAS or server, with your own AI key and budget in place of billing.

    In development
  4. 4

    Saving from Lingua Layer

    Save an expression with its sentence from the page you are reading, and see the ones you know marked on your next read.

    Planned
  5. 5

    Concepts inside the app

    Concepts, senses and fixed expressions live in a shared dictionary. What you saved, what you know and how you reviewed it stay in your own record and never write back into it.

    Being rebuilt

Next

  1. 6

    All eleven languages in the app

    After Japanese, Simplified Chinese and English, the rest of the eleven, then the next language families.

    Planned
  2. 7

    Knowing, judged sentence by sentence

    Sentence-level judgment of the words you know, and model routing, with the cost of every request shown and the default monthly cap in place.

    Being rebuilt
  3. 8

    Measuring the head start

    How much one language helps with the next, measured between related languages before any number is claimed.

    Being rebuilt
  4. 9

    Billing, after the free beta

    Only once the free beta shows the loop works: a cost on every request, caps, no automatic top-up, and export never blocked.

    Planned
  5. 10

    Video and audio

    Cut by utterance, sharing the same vocabulary and review state.

    Planned

Later

  1. 11

    Opening the core

    Licences chosen, the API and MCP documentation published, Anki files in and out.

    Intent
  2. 12

    Speaking and writing

    What you tried to say, the correction you kept, and whether you could use it on your own a week later. The ambition, not something that exists.

    Intent

When this page changes

This page changes when a piece changes: its label moves, and the moment is added to the timeline. Nothing moves on a schedule.

Questions about any piece, or collaboration [email protected]