Use case · offline and remote

Translate where there is no signal.

Lector runs on your own laptop. The reader, the on-device dictionary, cloze practice, and your vocabulary all work with the network off. Add a local model and you also get sentence translation and grammar help. No API key. No per-word bill. No signal.
Self-host, freeSet up a local model

Set it up while you have a connection. After that it needs none.

The honest split

What works with the network off.

Most of Lector never touches the network. Three features do, and one of them has a local answer. Here is the whole picture, so you can plan around it.

Works offline

  • The reader. Import an EPUB file, or paste text straight in.
  • Word lookup. The on-device dictionary answers from the language pack.
  • Cloze practice. Each pack ships a frequency-ordered sentence bank.
  • Your vocabulary, your streaks, and your statistics. Lector writes them to SQLite on your machine.
  • With a local model: sentence translation, phrase translation, the tutor, and journal correction.
  • Anki, both ways. Point the Lector add-on at your own server. AnkiConnect also works.
  • Speech for Esperanto. The container runs eSpeak NG, so it needs no key and no network.
  • Backups. Copy one directory to a second disk.

Needs a connection

  • The first install. Docker pulls the image once.
  • Language packs. Lector downloads a dictionary the first time you enable its language.
  • The model download. A model file is between about 3 GB and 8 GB.
  • Import from a web page, from YouTube, or from a podcast feed.
  • Google Cloud speech, which every pack except Esperanto uses for its best voice.
  • Claude, and any other cloud model.

Word lookup tries the on-device dictionary first. Only a miss goes to a model, so a local model answers the rest. Speech has a second route. When the call to Google Cloud fails, Lector uses the voice in your browser. A voice that your operating system installs speaks offline. Test your own browser with the network off, because some browser voices come from a server.

The setup

Three parts, all on your hardware.

Lector holds your books and your vocabulary. A language pack holds the dictionary and the sentence bank. A local model answers what the dictionary cannot. Each part lives on the same machine.

1. Run Lector on your own machine

One file for Docker Compose starts it. The default file also starts Ollama beside Lector, so the local model needs no separate install.

Installation guide

2. Install the language packs

While you still have a connection, enable each language in the picker. Settings marks each dictionary as installed at the end of the download. Dictionary files range from about 3 MB to about 280 MB.

Supported languages

3. Point Lector at the local model

Open Settings and choose the local provider. LM Studio and Ollama both work, and so does any server that speaks the OpenAI chat API.

Local model guide

A tip for a long trip. Set DICT_LANGS in your Compose file to name the languages you want, and Lector fetches each dictionary at boot. Use all to take every published language.

Who this suits

Built for a laptop a long way from a tower.

The same setup answers a common problem. You need a language, you have time, and you have no data connection to spend on it.

Field translators and linguists

You work with text far from a data connection. Lector keeps the reader, the dictionary, and the vocabulary record on your laptop. The text is not always yours to publish. With Lector, it never leaves the machine.

Missionaries and church workers

Language study in a remote posting cannot depend on a signal. Read at your own pace, tap the words you do not know, and practise them on the same laptop. Languages from common missionary fields are not supported yet.

Ask for your language

Aid and health workers

A deployment gives you weeks of dead time and no data connection. Load the books before you fly. Your reading and your review schedule continue with the network off.

Researchers, students, and crews at sea

A field station, a research vessel, or a long flight all give you time and no internet. Lector needs the laptop, and nothing else.

Read this first

Language coverage, without the sales pitch.

Lector ships 19 language packs. No Pacific Island language has a pack yet. There is no pack for Tok Pisin, Fijian, Samoan, Tongan, or Bislama. If you go to the Pacific for one of those languages, Lector gives you the reader and your vocabulary record, and no dictionary.
Other common fields are also open. There is no pack for Swahili, Thai, Khmer, Nepali, or Amharic, and none of them is on the roadmap yet. Hindi is on the roadmap, which helps in Fiji in part. Fiji Hindi is a separate variety, so treat that pack as a start and not an answer. Tagalog is on the roadmap for work in the Philippines.
Ask for the language you need with the form below. An unlisted language counts too.

Indonesian

The language of school and government in Papua and West Papua.

Pack detail

French

Official in New Caledonia, French Polynesia, and Wallis and Futuna. Also official in Vanuatu.

Pack detail

Spanish

Official on Easter Island, and across Latin America.

Pack detail

Mandarin

Spoken by Chinese communities across the region.

Pack detail

These four packs cover the languages of school and government in much of the region. See every pack for the rest.

Which language do you need in the field?

Name the language. We will email you once, on the day the pack is ready. No newsletter.

Your build order comes from this list. Want to talk it over? Join the Discord.

We keep your address for this one announcement. Read the privacy policy.

Before you travel

Do these eight things while you have a signal.

Every download happens once. Do each one at home, and test the whole loop with the network off before you pack.
  1. 1Start Lector once at home. Open it in a browser and read one page.
  2. 2Enable every language you need. Wait until Settings shows each dictionary as installed.
  3. 3Download the local model. Ask it for one translation, so you know the server answers.
  4. 4Import the books you plan to read. Paste any article you want to keep.
  5. 5Turn off the network. Tap five words. Practise five cloze sentences. Play one word aloud.
  6. 6Copy the data directory to a second disk. That copy is your backup.
  7. 7Check the memory. A 4-bit model of about 8B parameters needs 5 GB to 6 GB of RAM.
  8. 8Plan the power. A local model uses more battery than the reader alone.

The backup guide covers the data directory. One copy protects your whole library.

Straight answers

The questions that matter most.

For the way Lector teaches a language, read the method.
Does Lector work with no internet connection at all?

Yes, after you set it up. The reader, the on-device dictionary, cloze practice, and every statistic run on your machine. Add a local model and you also get sentence translation, grammar help, and journal correction. Setup itself needs a connection, because the image, the packs, and the model all download once.

My language has no pack. What do I get?

The reader works with any language. It tracks which words you know, and it keeps your reading progress. Without a pack there is no on-device dictionary and no cloze bank. A local model can still translate, but a small language tests its limits. Test the exact language on the exact model before you travel.

Which languages have a pack today?

Lector ships 19 packs: Mandarin, Spanish, Arabic, French, Portuguese, Russian, Indonesian, German, Turkish, Italian, Polish, Ukrainian, Dutch, Afrikaans, Czech, Swedish, Esperanto, Koine Greek, Latin. No Pacific Island language has a pack yet. Ask for the one you need with the form on this page, and see the languages page for the detail of each pack.

Do I need a GPU?

No. A 4-bit model of about 8B parameters runs on a laptop with 5 GB to 6 GB of free RAM. A GPU makes each answer faster, and it changes nothing else. You can also skip the model. Word lookup then comes from the on-device dictionary alone.

Is a local model good enough for translation?

A capable local model holds up well, even for a language with little data. In our own test of Afrikaans to English, a 12B model scored 82.7 chrF++ on Tatoeba, ahead of Claude Opus and GPT-4o. Two cloud models beat it, by less than a point. Pick a general instruction-tuned model of 8B parameters or more. Read the full numbers in the translation eval.

Does my text leave the laptop?

With the local provider, no. Lector calls your own model, and the text stays on your hardware. The cloud provider is the exception, because Claude runs on Anthropic servers. See the privacy page.

Can Lector speak a word offline?

In part. Esperanto speaks through eSpeak NG, which runs inside the Lector container and needs no network. Every other pack calls Google Cloud for its best voice, and that call needs a connection. When it fails, Lector uses the voice in your browser instead. A voice that your operating system installs speaks offline, so test your own browser with the network off before you travel.

Does Anki sync work offline?

Yes, and both routes stay on your machine. The Lector add-on for Anki is the route we recommend, because it syncs both ways. Point its api_url at your own server, such as http://localhost:3457. AnkiConnect also works, because the browser talks to Anki on the same machine.

Can a team share one laptop?

Yes. Lector serves the web app and the API over the local network, so a phone or a tablet on the same network can open it. Set API_URL to the address the browser uses. The installation guide covers it. Give each person a separate library with multi-user accounts.

Can I install it with no connection?

No. Docker pulls the image, Lector downloads the dictionaries, and the model server downloads the model. Do all three before you leave. After that, Lector needs no connection.

What does it cost?

If you host it yourself, nothing. Lector is open source, and there are no feature gates. A local model costs no money per word. Lector Cloud starts at $5 a month, and it needs a connection, so it is the wrong choice for the field.

Set it up before you go.

Lector is free to self-host, and it is open source. Install it tonight, load your packs, and test it with the network off.

Read the installation guideAsk on Discord

Is there no pack for your language? Write to us at [email protected].