The Koboshi blog

You can now learn Hindi, Urdu, and Bengali from English Roman letters

Sun, 04 Oct 2026 00:00:00 GMT

Koboshi now supports Hindi, Urdu, and Bengali answers in English letters. In the quiz, you can type a Hindi, Urdu, or Bengali word in roman letters, the way many of us text, and it counts. Every Hindi, Urdu, and Bengali word you save also keeps its romanized spelling.

Why are English letters needed for Hindi, Urdu, and Bengali speakers and learners?

We know you could perhaps write, but do not bother to write it, especially on digital devices.

For Urdu, a 2025 paper mentions Roman Urdu as “the dominant form of informal communication on platforms like Instagram, WhatsApp, and social media”, and its popularity is reasoned as “the lack of easy-to-use Urdu keyboards and familiarity with Latin characters” (Butt, Varanasi and Neumann, RANLP 2025). For Hindi, a 2014 study of 1,062 Hindi-English comments on the public Facebook pages of Amitabh Bachchan, Shahrukh Khan, Narendra Modi and BBC Hindi, collected in October 2013, found 93.17% written in Roman script and 2.93% in Devanagari (Vyas et al., EMNLP 2014).

The study is not the latest piece, but the trend should be almost the same.

Hence, the learning experience of Hindi, Urdu, and Bengali should not be stressful for their speakers, especially if they are heritage speakers who could speak already. Since at Koboshi, you could refresh your heritage language while learning the new language, the process needs to be stress-free enough to focus on the main learning language.

If you are learning Hindi, Urdu, and Bengali mainly as a learning language, then you should try to answer using their native fonts. This upgrade is just an easy and a bit lazy way for the natives to get around.

Which spellings count?

Each line counts as one word, whichever spelling you type:

mohabbat or muhabbat (محبت, मोहब्बत, love)
zaroor, zarur or zaror (ضرور, ज़रूर, of course)
theek, thik or thek (ٹھیک, ठीक, fine)
bhalo, bhaalo or valo (ভালো, good)
khub or khoob (খুব, very)

The quiz would be extra lenient due to variations in how people express themselves in roman letters.

Why is Koboshi a great app for Hindi, Urdu, and Bengali speakers?

If you have roots in the Subcontinent (India, Pakistan, and Bangladesh), then you are most likely at least bilingual or more. English sits mostly with the educated and urban families. In India’s 2005 national household survey, about one adult in five could speak at least a little English, but 88.5% of college graduates could, against almost none of the adults with no schooling, and 34.8% of city dwellers against 14.1% of people in villages (Azam, Chin and Prakash, using the India Human Development Survey 2005). Pakistan has no national count of English speakers, but two numbers can be counted: in urban Punjab, 47% of children aged 6 to 16 go to private schools (ASER Pakistan 2023, Punjab), and UNESCO counts 114,849 Pakistani university students abroad in 2023, up from 24,601 in 2005 (UNESCO Institute for Statistics). In Bangladesh, English-medium schooling is a small and expensive slice: 148 of more than 21,000 registered secondary schools (BANBEIS, reported by The Business Standard, 2026). Also given huge social inequality in these regions, the chance is that most proactive learners would fall into the multilingual side.

So, if you are learning a new language, then you would still want your native language to be shown so you could relate better and study well.

Also, did you know that Urdu, Hindi, and Bengali speakers who grew up with English at school are likely weaker in their native tongue than people who are truly monolingual, thanks to the society tolerant to English where people revert to English words in their daily conversations? The research points the same way. Children who grow up with two languages know fewer words in each single language than monolingual children: in a pooled study of 1,738 children aged 3 to 10, those who spoke another language at home (Hindi, Punjabi and Urdu among them) and English at school scored 96.3 against 106.8 on an English vocabulary test (Bialystok et al., 2010). A 2024 study of 743 toddlers in Montreal found the same in each single language, while both languages together came to more words than one (Byers-Heinlein et al., 2024). Linguists describe a language used mainly at home as lagging behind the majority language in vocabulary (Montrul, 2012).

Online, that mixing can be counted. In the 2014 Hindi Facebook study, about one in four comments written mainly in Hindi (93 of 398, 23.37%) slipped English words into Hindi sentences (Vyas et al., EMNLP 2014). In 51,277,891 words of the Hindi newspaper Dainik Jagran, “party” appeared “over 67K” times, while “the counts for the equivalent Hindi forms are relatively low” (Bali et al., 2014). For Bengali, about one in five posts (21.05% of 2,309) in campus Facebook groups at Jadavpur University mixed Bengali and English (Das and Gambäck, ICON 2014). And the 2025 Roman Urdu paper counts “frequent code-switching with English” among the complexities of Roman Urdu (Butt, Varanasi and Neumann, RANLP 2025).

How to save words to Koboshi from ChatGPT and Claude

Mon, 28 Sep 2026 00:00:00 GMT

Ask ChatGPT or Claude to save any word to Koboshi, and the word lands in your Koboshi vocabulary.

ChatGPT users: jump to ChatGPT. Claude users: jump to Claude.

Set it up once per app. The address is https://koboshi.app/mcp

ChatGPT and Claude change their screens often, so what you see may differ a little from these pictures. Checked on 2026-09-29.

ChatGPT

Open Settings → Plugins.
Press Browse directory.
Press Add → New Plugin.
Fill in Name Koboshi, the address https://koboshi.app/mcp, Authentication OAuth, tick I understand and want to continue, and press Create.
Sign in to Koboshi and press Approve. The first time, ChatGPT also asks the month and year you were born, the same question the Koboshi website asks.
Press Try in chat.
In any chat, type @Koboshi and ask it to save a word.

Check: if you are unsure it worked, open Koboshi and look for the word.

Claude

Open Settings → Connectors.
Press Add.
Type Koboshi as the name, paste https://koboshi.app/mcp, and press Continue.
Press Connect next to Koboshi, sign in, then press Approve. The first time, Claude also asks the month and year you were born, the same question the Koboshi website asks.
In a chat, press + → Connectors. Koboshi is now on the list: switch it on.

Check: ask Claude to save any word to Koboshi.

Using Claude Code or Codex? Connect them here. Steps checked on 2026-09-29.

How to save words to Koboshi from Claude Code and Codex

Mon, 28 Sep 2026 00:00:00 GMT

Ask Claude Code or Codex to save any word to Koboshi, and the word lands in your Koboshi vocabulary.

Two commands in a terminal and one sign-in per app. These tools change often, so check their own docs if a command stops working. Checked on 2026-09-29.

Claude Code

Run these two commands. The second opens Koboshi: sign in and press Approve. The first time, Claude Code also asks the month and year you were born, the same question the Koboshi website asks.

claude mcp add --scope user --transport http koboshi https://koboshi.app/mcp

claude mcp login koboshi

Check: this shows koboshi as ✔ Connected.

claude mcp list

If claude mcp login is not found, start Claude Code, type /mcp, pick koboshi and sign in.

Codex

Run these two commands. The second opens Koboshi: sign in and press Approve. The first time, Codex also asks the month and year you were born, the same question the Koboshi website asks.

codex mcp add koboshi --url https://koboshi.app/mcp

codex mcp login koboshi

Check: koboshi is on this list.

codex mcp list

One setup covers the Codex CLI, the IDE extension and the ChatGPT desktop app.

Using ChatGPT or Claude? Connect them here.

Why I built Koboshi: a vocabulary app for people who lived using more than one language

Mon, 17 Aug 2026 00:00:00 GMT

Koboshi is a vocabulary app for people who live in more than one language. You look up a word once, and Koboshi shows you what it means across the languages you already know, saves it, and brings it back for review right before you would forget it, on a spaced-repetition schedule (FSRS, the same family of scheduling that serious flashcard systems use). That is the whole product in one sentence. This post is about why it exists, and the story is quite personal.

The moment that made it click

“Ah, THAT word” moment does not always come from the same language.

When I look up a Spanish word and see its meaning, sometimes the click comes from the English meaning. Sometimes English does nothing for me and the Urdu meaning lands it instantly. Sometimes it is the Japanese. The language that unlocks a word is not predictable, so an app that shows you the meaning in one language is not enough for me.

Koboshi shows the meaning across your languages, side by side. And that produces a quiet second effect I did not fully appreciate until I lived with it: while you study the new language, your weaker languages get exercised too. Every lookup is a small rep for them. One stone, two birds: 一石二鳥, as we say in Japanese. The research points the same way, from several directions at once: a language you grew up speaking is dormant, not lost, and actively recalling it is what wakes it up.

The heritage speaker part is personal

I grew up in Pakistan. I speak Urdu like a native, because I am one: the rhythm, the accent, the jokes, all of it. But if I am honest about my Urdu vocabulary, it sits around A2, maybe B1. Fluent mouth, thin dictionary.

Part of that is just how Urdu works day to day: we swap English words into Urdu sentences constantly and they fit, so you can be a full member of the society without ever needing the harder native vocabulary. Plenty of educated people around me had the same tendency, just with a bit more schooling behind their word choices than mine.

That combination, sounding native while missing the vocab, has a name: being a heritage speaker. Koboshi is built for heritage speakers as much as for classroom learners.

Who this is actually for

If you already hold two or more languages and are adding another, Koboshi was built around your exact situation: the new language gets the serious spaced-repetition treatment, and your existing languages ride along, getting stronger instead of rusting.

And if one of your languages is a heritage language you would like to finally give the vocabulary it deserves, this is for you. Surprise your friends and family in other parts of the world in six months.

However, even if you do not have a heritage language, you could still use Koboshi, and it should be nearly as helpful as it is for the heritage speakers. You still get to learn a new language while strengthening the other you already know. Even if you are learning the first foreign language, it is still useful for you to use it as a learning engine and your personal dictionary.

The real origin story: a ChatGPT project that kept forgetting

Before Koboshi existed, my system was a ChatGPT project. Every new word I met went into the chat. Most of them came from Duolingo: it tracks the words its lessons introduce, but the listening exercises kept handing me words that never showed up in its tracking, and I wanted to keep those.

It worked, for about two or three months at a time. Then the chat would fill up. A conversation has a memory limit (the context window), and when it fills, the oldest content quietly falls out. Words I had carefully saved were just gone. So I would start the next chat in the same project and begin the pile again.

ChatGPT will happily offer to export everything into a spreadsheet, and that fixes nothing: the memory limit is still there, and no one in 2026 is learning vocabulary out of a spreadsheet. What I needed was not storage. I needed to record a word in the same motion as meeting it, and to be able to learn it in the same place, with review that happens by schedule instead of by willpower.

Building that had recently become possible for one person. So I built it. Koboshi is the tool I wished existed during those three-month cycles: look it up, see it across your languages, it saves itself, and it comes back to quiz you right before you would have lost it. The name comes from 起き上がり小法師, the little Japanese doll that rights itself every time it falls, and from 取りこぼし, the things you let slip while trying to hold on. Words you looked up and lost. That is exactly what this app rescues. Cool Gaijin stuff right?

Where it goes from here

Koboshi is young, and it gets better every week. If the situation I described is yours, try it with a handful of words you met today.

Why you understand Spanish but can't speak it: the lexical retrieval lag problem

Mon, 01 Jun 2026 00:00:00 GMT

You can follow the conversation. You catch the jokes. But when it is your turn, the word you need arrives a beat too late, and the moment is gone. The reason is simple and fixable: understanding and speaking run on two different skills. Recognizing a word when you hear or read it is one skill. Pulling that word out of memory on demand, fast enough for a live sentence, is a separate skill called retrieval, and most studying never trains it. Conversation runs almost entirely on retrieval, and retrieval is trainable on its own, without a conversation partner, in minutes a day.

Updated August 2026: rewritten around the research on retrieval practice, typing versus speaking, and word combinations, with sources linked where each claim is made.

Recognition is not retrieval

Every learner’s passive vocabulary (words you understand) runs far ahead of their active vocabulary (words you can produce). That lag is not a personal failing; it is the documented norm in second-language research (Laufer, 1998, Language Learning). Understanding gives you generous time: the word is right there in the sentence, and context does half the work. Speaking gives you no time at all.

A half-known word costs you a beat. By the time it arrives, the next sentence is already gone.

So the gap you feel in Spanish is not evidence that you learned the wrong words. It is evidence that you trained recognition and never trained retrieval.

What actually trains retrieval

The research here is unusually one-sided. Actively recalling a word from memory beats re-reading or re-viewing it by about half a standard deviation of extra retention, measured across 159 separate comparisons (Rowland, 2014, Psychological Bulletin; the effect size is g = 0.50, where g is a standard unit for sizing the difference between two methods: 0.2 counts as small, 0.5 as medium, 0.8 as large, so this is a solid, clearly noticeable advantage). The classic experiment used foreign-language vocabulary specifically: repeated self-testing produced large gains a week later, repeated studying produced almost none, and learners could not feel the difference while it was happening (Karpicke and Roediger, 2008, Science).

Three details in that literature matter for the speaking problem:

Direction matters. Practicing from the meaning to the word builds the ability to produce the word; going only from the word to the meaning does not transfer well to production (Terai, Yamashita and Pasich, 2021, Studies in Second Language Acquisition).
Retrieval practice builds speed, not just accuracy. Words you have actively recalled come back faster next time (van den Broek and colleagues, 2014, Memory), and speed is the point: fast, automatic access to vocabulary predicts speaking fluency, while the sheer size of your vocabulary does not (Takizawa and colleagues, 2025, Applied Linguistics, a study of 210 learners).
Spacing beats cramming. Spreading retrievals out over days beats massing them into one session, across 48 experiments (Kim and Webb, 2022, Language Learning).

Do you have to speak to close the gap?

Honest answer: partly, and it helps to know exactly which part.

The words themselves transfer. In the cleanest direct experiment, learners who studied words by writing scored 68 percent when tested in writing and 58 percent when tested in speech, and even that gap disappeared within a week; the researchers went in expecting words to work best in the mode they were learned in, and their data rejected that prediction (Candry, Deconinck and Eyckmans, 2018, Journal of the European Second Language Association). Typed recall genuinely builds the words you will later speak.

What typing cannot do is automatize the act of speaking itself. Skills automatize separately: comprehension practice speeds up comprehension, production practice speeds up production, and neither substitutes for the other (DeKeyser, 1997). Speaking fluency in training studies improves through spoken repetition under time pressure, such as telling the same story in four, then three, then two minutes (de Jong and Perfetti, 2011, Language Learning).

The clean way to say it: typed retrieval builds and speeds the engine; speaking connects that engine to your mouth. A typed quiz alone will not make you a fluent speaker, and you should distrust any app that claims otherwise. The bridge costs nothing: say your answer out loud as you type it (saying items aloud adds a reliable memory edge over silent study, shown durable for foreign vocabulary: Icht and Mama, 2019, Language Teaching Research), and retell things you can already say, a little faster each time.

Learn combinations, not only words

In Spanish you do not really “take” a walk, you give one: dar un paseo. Knowing paseo is not the same as being able to say dar un paseo when the moment comes.

The research on multi-word combinations is honest in both directions. They are not easier to memorize; in direct comparisons, combinations carried a heavier learning burden than single words (Peters, 2016, Language Teaching Research). But they are what fluent speech is made of: natural-sounding speakers assemble sentences largely from prefabricated chunks (Pawley and Syder, 1983), and learners’ chunk knowledge lags badly, sitting below 50 percent even for combinations built from the 1,000 most frequent word families (Nguyen and Webb, 2017, Language Teaching Research). Deliberately studying combinations closes that gap, with large gains in a 2024 meta-analysis of 17 studies (Li and Lei, 2024, IRAL; d = 1.415, where d is another effect-size unit and anything past 0.8 counts as large, though this one pools varied study designs, so read it as “works well”, not as an exact constant).

There is even a memory trick hiding in the combination: attaching a new word to a word you already know gives it a retrieval cue you will actually meet in real sentences (Kasahara, 2011, System). So when a word arrives inside a combination, save the combination.

What this looks like day to day

The routine the evidence supports is small and specific:

See the meaning, produce the word from memory. Meaning to word, not the reverse.
Say it out loud while you type it.
Space the repetitions out; let the misses come back sooner than the hits.
Watch your speed improve, not just your accuracy.
When a word shows up inside a combination, keep the combination.

This pattern is what Koboshi’s Hard quiz is built on: typed recall from the meaning, spaced by a scheduler, and, since August 2026, timed, because response speed is the number the fluency research says actually matters. Phrases and multi-word expressions can be saved alongside single words; a dedicated combination exercise is something we may add in the future.

Whichever tool you use, the shape of the fix is the same: a few minutes of producing words from memory, daily, beats hours of re-reading, weekly. The gap between understanding and speaking is normal, it is measurable, and it closes from the retrieval side. The words stop arriving late.

For years, when someone has asked me whether I speak Spanish, my answer has not been “hablo un poco de español”. It has been “entiendo más de lo que puedo hablar”: I understand more than I can speak. Every learner carries a version of that sentence. Train the retrieval side, and someday we will not need to say it.