mndict · experimental AI dictionary for Android specimen — ← mnsky

petrichor

/ ˈpɛtrɪkɔː / noun a prepared mndict entry

The smell of dry earth when the first rain lands on it.

mndict is an experimental Android dictionary. Look up words and phrases, keep a collection, and practice what you saved. Bring your own model provider.

Prepared examples written for this page. No request leaves your browser.

Nothing looked up yet. Type a word, or pick one from the list beside this column. Six of them have entries here; anything else returns the app's honest answer.

Look up anything, not just headwords

Words, phrases, idioms and short sentences — explained in the shape you ask for, instead of whatever a chat thread happens to produce. A three-month-old conversation never becomes the index you have to scroll.

That is the whole premise: the output format is a contract, not a mood.

One collection, readable without the network

Auto-save is on by default. Saved entries open without another provider request, so the collection is still there on a plane with no signal. Search, Words and Practice are independent sections rather than one endless list.

  • Search — look something up.
  • Words — everything you kept, organised.
  • Practice — drill the meanings you saved.

The shape of an entry

Every lookup comes back as plain text in the same structure. Definition and usage are required. Examples are natural sentences, each showing a different sense. Synonyms and antonyms are real substitutions or the word None — and if the app is not sure the entry exists, it says exactly that instead of inventing one.

DefinitionTwo or three sentences on the core meaning, connotation included.
UsageRegister and context, and what a learner would get wrong.
ExamplesFive sentences, never the same pattern twice.
SynonymsWords a native speaker would actually use instead.
AntonymsDirect opposites only.
NoneThe honest answer when there is nothing to give.

Rearrange it like furniture. The Prompt Builder toggles sections, drags blocks into your order, and sets how much each one says — brief, normal or detailed. Translation and etymology are there if you switch them on.

Bring your own model

The app does not ship a model and does not sell one. You point it at a provider, paste a key, and the requests are yours. A local Ollama endpoint works the same way as a hosted one, if you have the hardware for it.

An earlier experiment ran small models on the phone itself. A roughly 2.7B-parameter quantised model turned out to be both slower and weaker than a provider call, and the local-inference path was retired. That history is why the current app is provider-backed.

Thirteen providers, a local Ollama slot, and one custom slot for anything else that speaks the OpenAI wire:

GeminiOpenAIAnthropicGroqDeepSeek MistralxAITogetherCerebrasFireworks OpenRouterNVIDIA NIMOllama (local)Custom endpoint

Getting it

There is no public repository for mndict yet, so there is nothing to download from this page. The app runs on an Android arm64 device and is still being tinkered with — if you want to try it, ask and it can be arranged.

Explanations come from a language model and can be wrong. Treat it as a reading aid, not a reference work, and double-check anything that matters.

Made by
Tyan
Runs on
Android, arm64
Built with
Flutter
Interface
Prism — small lexical type on a continuous ambient ground
Providers
Your key, your bill
Status
Experimental