For authors, translators and publishers

Translate the book.Keep the thread.

EPUB, PDF or chapters: Libris Translate translates the whole book, with a memory for its names and your chosen AI model. You keep the text and the editorial decisions.

Your server. Your model. Your files.

Libris Translate showing a passage of an English book next to its French translation, under the book's progress from import to export.
The source and its translation, in the Libris editor.

Read the result.

Two Alice books translated end to end. The files, reports and limitations are public, including the declared human edit.

Excerpts and downloadable books

Lewis Carroll · English → French · Libris 0.25.0

Source text

The way Dinah washed her children’s faces was this: first she held the poor thing down by its ear with one paw, and then with the other paw she rubbed its face all over, the wrong way, beginning at the nose […]

Translation

Voici comment Dinah lavait la figure de ses petits : elle commençait par maintenir le pauvre petit par l’oreille avec une patte, puis, de l’autre, lui frottait toute la figure à rebrousse-poil, en partant du nez. […]
Actual excerpt from Through the Looking-Glass. Omissions marked […].

What sentence-by-sentence translation loses

Translate a novel one passage at a time and you lose track of who is speaking, what a place is called and which word was chosen three chapters back. Libris Translate analyses the book first and writes what it learns into a memory that every later passage reads.

A memory that follows the series

Before translating, Libris Translate writes a Book Bible for each volume: characters, relations, places, terms. A series adds a Series Bible and a glossary shared by all its volumes, and several series set in one universe can share a glossary too. A volume only learns from the volumes before it, so volume 3 cannot give away a twist from volume 5.

Example written for this page

Tide Keeper → Gardien des marées Glossary term, locked · since volume 1

  1. Volume 1 The Tide Keeper never left the lighthouse. Le Gardien des marées ne quittait jamais le phare.
  2. Volume 7 Even now, the Tide Keeper waited for Mara. Même à présent, le Gardien des marées attendait Mara.

An autopilot for the whole book

Analysis, translation, review, a final review and arbitration of the remaining doubts run one after another, with nothing for you to approve. A failed passage is retried with another strategy. If a provider goes down, Libris Translate waits, retries, then moves to the next provider in the fallback chain; near the book's budget it switches to a cheaper one. Each decision is written to the report.

  • At most three convergence rounds by default (configurable from 1 to 10). The report says whether the result is ready, ready with recorded details, or blocked by untranslated passages, unresolved errors or checks without a valid result.
  • A passage no model could translate keeps its original text and its reason. A provisional export can include it; that does not make the book ready.
  • A job interrupted by a restart resumes from its last checkpoint.

How it works

Four steps from a file to a translated book. Only the first and the last one need you.

  1. Add a book

    Drop an EPUB, chapter files, or one long file that Libris Translate splits at its chapter headings. The import assistant suggests the series, the volume number and the place of each prologue, interlude or split chapter, and says why.

  2. Let it run

    Libris Translate analyses the book several passages at a time, builds its memory, translates, reviews and settles open questions on its own.

  3. Look in if you want

    Follow the progress, open the passages with the lowest scores, fix a sentence or lock a term. Nothing waits for your approval.

  4. Download

    An EPUB checked with EPUBCheck, a bilingual EPUB for proofreading, a Word file, or the chapters as text, Markdown or a ZIP, with the autopilot's report.

Recorded with Libris 0.23.0. Follow the written guides for version 0.29.0; some screens have changed.

Libris Translate in 43 seconds · 43 s A real recording, without sound: Libris Translate 0.23.0, set up with the official installer, translates three tales from Alphonse Daudet’s “Lettres de mon moulin” (public domain) from French into English with GPT-6 Luna. The long steps are sped up.
What the video shows, scene by scene
  • The libris-translate.com website, then the Install button.
  • A Linux server: the official install command downloads Libris Translate and starts it, up to the message “Libris 0.23.0 is ready at http://localhost:8088”.
  • I. Add a book: the French EPUB of “Lettres de mon moulin” (5,964 words, 18 passages) is imported as a standalone volume, from French into English, with the GPT-6 Luna model of a ChatGPT account.
  • II. Let it run: the analysis, the translation and the review, sped up, then the glossary the analysis built: Jean de Milan becomes John of Milan, mas farmhouse, miarro farm boy, meule millstone and parc sheepfold.
  • III. Take a look if you like: the French text and its English translation side by side, passage by passage.
  • IV. Download: the translated EPUB, checked by EPUBCheck without a single error, open on an e-reader at “The Secret of Master Cornille”.
  • “Your next series, translated on your server.” 30-day free trial at libris-translate.com.

Who it is for

  • Independent series authors

    Your saga in another language, with the same names from volume 1 to volume 10, for every bookshop.

  • Translators

    Start from a complete draft and a score for every passage; your corrections are never overwritten.

  • Small presses and digital publishers

    A consistent first draft of the whole series, on your server, and the map of the passages to review first.

  • Webnovel and series readers

    Follow a series as it is published. New chapters are translated on their own, with the earlier ones as context, and a name stays the same from volume 1 to volume 12.

  • Self-hosters

    Runs in Docker on a Linux server: one Compose stack with PostgreSQL 17, and a documented nightly backup that runs without stopping anything. You decide which model providers see the text.

  • Developers

    A REST API described in OpenAPI 3.1, signed webhooks, and an MCP server for coding agents. Send an EPUB or JSON chapters, get the translation back with a report.

Features

  • Series and memory

    Series with numbered volumes, or an open-ended stream of webnovel chapters. A Book Bible per volume, a Series Bible per series, character identities and relations, and glossaries per book, per series or for a whole universe. Change a term at chapter 300 and Libris Translate counts where the old word still appears, then offers to replace it without calling a model.

  • Webnovels still being published

    New chapters join their volume and only they are translated, with the earlier chapters as context. Duplicates are skipped. Libris Translate can also watch an RSS or Atom feed that carries the chapter text, every six hours by default; it does not scrape web pages.

  • Autopilot

    Failed passages are retried, providers fall back during outages, remaining doubts go to AI arbitration, and every decision lands in a log you can filter by stage or passage. Analysis works on several passages at once, each one seeing only the text before it.

  • Formats kept intact

    EPUB 2 and 3 in, up to 256 MB per file by default, with layout, styles and images preserved, and EPUBCheck 5.3.0 on the way out. Chapters as TXT, Markdown, HTML or DOCX, text PDFs and scans with local OCR, and JSON through the API. PDF layout and images are not reproduced; there is no PDF export. A bilingual EPUB, interleaved or side by side, and a Word file with your corrections as tracked changes and the model's doubts as comments.

  • Where to proofread first

    Every passage gets a score from 0 to 100, computed from the alerts, critiques and doubts Libris Translate already records, with no extra model call. Four bands: good from 85, fair from 70, weak from 50, poor below. Under 70, Libris proposes a review, the 20 weakest passages first. Repeated turns of phrase are flagged for a look in context, not treated as proof of a flattened style.

  • Your corrections win

    A side-by-side editor with version history, AI suggestions you accept or refuse, locked terms, and glossary imports previewed before they apply. The autopilot never rewrites a passage you corrected or validated.

  • Your model, your costs

    Any OpenAI-compatible server, local or hosted, the OpenAI and Anthropic APIs, and an optional Codex/ChatGPT bridge. Every paid operation shows an estimate first. Budgets can be set per book and per API token: at 90 %, Libris Translate moves to a cheaper provider if the fallback chain has one, otherwise it pauses the job. A price list in another currency is converted once, at the European Central Bank's daily rate.

  • Half price for work that can wait

    OpenAI and Anthropic charge half for batch work they may take up to 24 hours to process, which suits a book translated overnight. Batch mode is turned on per provider, for long jobs; it is not available through a ChatGPT account. Libris Translate records each submission before sending it, collects batches again after a restart, and books the price actually paid. The 24 hours are the provider's processing window, not a delivery time.

  • Accounts and sign-in

    Local passwords with authenticator codes, WebAuthn security keys and ten recovery codes; OpenID Connect single sign-on; LDAP or Active Directory. Administrators manage local accounts in Libris Translate; SSO and directory identities stay with their provider.

  • Shared by a team

    Waiting jobs start by priority, then take turns between accounts and API tokens, so one large backlog does not hold everyone else up. An administrator gives each account a share of the monthly word quota, and each member can add model providers with their own keys, which nobody else sees.

  • A proofreader without an account

    With Studio and Pro, a review link opens one book, read-only, for 30 days by default, longer or with no end date if you choose, and you can revoke it at any time. It uses no account seat and sets no cookie. The source text is shown only if you chose to include it when creating the link.

  • Nextcloud and WebDAV

    Each account can browse approved Nextcloud, ownCloud, NAS or WebDAV servers from the import assistant, and send finished books back in any export format, by hand or automatically as each one finishes. Credentials are encrypted, connections are private to the account, and the administrator lists the hosts and private networks that may be reached.

  • Straight into your library

    A finished EPUB can go into a folder watched by Calibre-Web, Komga or Kavita, by hand or automatically. A failed delivery is retried without retranslating anything, and series metadata carries the series name and volume number so the books arrive in order.

  • Webnovel exports

    Libris Translate builds an EPUB for a serial that never had one: one page per chapter, a table of contents, EPUBCheck. Export a range, such as the last fifty chapters, or a single chapter as text or Markdown. A chapter that is only partly translated is refused rather than mixed with the original.

  • A series in one archive

    Export a series, all its volumes including archived ones, and its editorial memory into one checked archive, then restore it elsewhere without reusing identifiers or carrying provider secrets. Special volumes such as DX1, EX or LN 14+ keep their label and their place in reading order.

  • Email when it matters

    Libris Translate writes to you when a book is finished, when a job has stopped and needs a decision, when the quota cycle reaches 80 % and when the licence has not been renewed for 24 hours, never when a job merely starts. Each type can be turned off, and nothing is sent until you enter an address.

  • Automation API

    A versioned REST API with scoped tokens, idempotent requests and signed webhooks, described in OpenAPI 3.1, with an example Python client. Send a book or new chapters, get the translation and its report back.

  • MCP for coding agents

    Libris Translate answers the Model Context Protocol at POST /mcp. opencode, Claude Code, Codex and other clients can read the library, set up a book, edit the glossary, start and steer jobs and read the result, with the same accounts, rules and audit log as the interface. A token only gets the scopes you tick.

  • Books stay on your server

    Books and translations are stored on your server. Model calls send passages with their context, including glossary and narrative memory, to the providers you configure. Optional memory, search, mail and delivery services receive the data needed for the features you enable.

  • Documentation for your version

    The API reference and the changelog ship inside each image and open from within Libris Translate, so they describe the version you run, not the latest one on this site.

  • Interface in English and French, light and dark themes or the device setting, from a phone to a wide screen.

    Automation API

Pricing

Libris Translate runs on your hardware. The personal plan is a one-off purchase for non-commercial use, Studio is paid monthly or yearly and Pro is quoted. There is no account to create on this site, and none of your text passes through us.

Before buying, see what it produced on two real books

Studio

29 € per month, or 290 € a year

Commercial use: a translator working on a long series, with a few people reading along.

  • 3 000 000 words per month
  • Five machines, three accounts
  • Books shared between accounts, each person correcting their own
  • Automation API (REST, OpenAPI 3.1) and MCP server
  • Everything in Personal, for commercial use
  • Email support, first reply within 2 business days (indicative)
Buy a licence

Pro

On request by subscription

A team, a publisher, or an installation that serves many people.

  • Monthly words set in your quote
  • Several machines
  • Several accounts sharing books
  • Automation API (REST, OpenAPI 3.1) and MCP server
  • Everything in Studio
  • Email support, first reply within 1 business day (indicative)

You can start with the trial: the form below sends a confirmation link by email, then your key once you confirm your address. To buy, write to us and we reply with your licence.

How the licence works

  • A book is counted as soon as you add it, all of its words, even if you delete it without translating it. The quota counts the words of the source text; in Chinese and Japanese, two characters count as one word. Translating the book, retranslating it or correcting a passage adds no source words to the licence quota; model calls are still billed separately by the provider. Comparing several AI providers on the same passages is counted on top.
  • The quota renews every month on the date you bought your licence (or converted your trial), not on the first of the month.
  • A book is accepted as long as it fits in what is left of the cycle, plus a margin of 20 000 words; otherwise it is refused and nothing is counted. Words counted past the quota are taken off the next cycle.
  • You move machines yourself, from the licence screen. Releasing a machine frees its slot at once. Activate one more than your plan allows and the oldest one is released, provided it has held its place for at least seven days.
  • If our licence server is down, you keep working. Your installation holds a signed authorisation valid for 72 hours and renews it every hour.
  • Personal receives the new versions published in the twelve months after the purchase. After that, the software keeps working, security and compliance fixes are provided until 24 months after the purchase, and renewing the new versions is optional. Studio and Pro receive every version while the subscription runs. If the publisher ceases trading, every valid licence receives a certificate that works offline for good. See the terms
  • Personal is for non-commercial use: translating for yourself, for your own reading or as a hobby. Paid work, a translation you sell or publish for money, and use by a company or an organisation for its activity need Studio or Pro. See the terms
  • Your translations remain yours after the licence ends: you can still read, export and delete them.
  • The model is not included. You use your own, hosted or local, and pay its provider directly. Libris Translate shows the estimated cost before each paid operation. What a book costs with each model

Try it for 30 days

A trial key, sent to your address as soon as you confirm it: 100 000 words a month, one account, two machines, every format and provider, and the full autopilot. The automation API, the MCP server and book sharing (members and review links) come with Studio and Pro.

One trial per address, for life, and addresses with a “+” in them are not accepted. No payment card. We write to you about your trial (its key, its quota, its end) and, only if you tick the box, up to four tips to get started. Never advertising; unsubscribe in one click. After confirmation, the key email also holds the credentials to download Libris Translate and a link to the installation guide.

Your address is used to confirm it, to send the key and to grant only one trial, and for the tips if you ask for them. Privacy policy

licences@libris-translate.com

Where Libris Translate runs

Today, on your own server. A hosted version is planned but not available yet.

Available now

On your server

A Docker Compose stack on your Linux server. Model calls send passages and their context to your configured providers. Optional memory, search, mail and delivery services receive the data their enabled features need. The licence server receives activation and usage checks.

Go to the install steps
Planned

Hosted

Libris Translate, run by us for teams that would rather not operate a server. It does not exist yet. Leave your address and we will write to you once, when it opens.

Tell me when it opens

Install Libris Translate

You need a Linux (amd64) server with Docker Engine and Docker Compose v2, access to a language model, and the read-only registry credentials that come with your licence. The Libris Translate server needs no GPU when the model runs elsewhere.

Show installation commands
  1. Your licence email holds three things: a licence key, a registry username and a read-only registry token. Log Docker in as root first, since the installer also runs with sudo:

    bash
    sudo docker login registry.libris-translate.com

    At Username, type the registry username from the email. At Password, paste the registry token, not the Libris Translate licence key; nothing is shown while you paste. Docker answers Login Succeeded.

    Then install Libris Translate. The command pulls registry.libris-translate.com/libris/libris:latest; you need neither Git nor an archive:

    bash
    curl --fail --location --silent --show-error https://libris-translate.com/install.sh | sudo bash

    The installer takes the Compose definition from the image it is about to start, pins that image to its digest, creates /opt/libris/.env with random secrets, then starts PostgreSQL, the migration, the API and the worker. Running it again updates to latest, keeps the secrets, and only replaces Compose if you have not edited it. To install this exact release instead of latest: curl --fail --location --silent --show-error https://libris-translate.com/install.sh | sudo env LIBRIS_TAG=0.29.0 bash.

  2. Open http://localhost:8088 on the same machine and sign in as admin; the generated password is in /opt/libris/.env. Then go to Settings › Licence and enter the licence key from your licence email under Licence key:

    bash
    sudo grep '^BOOTSTRAP_' /opt/libris/.env

    On a remote server, open an SSH tunnel first: ssh -L 8088:127.0.0.1:8088 you@your-server. For network or HTTPS access, see Open Libris to your network in the Docker guide.

Already running Libris Translate? Pause the books in progress, back up the database, the books and the configuration together, then run the same command: it moves to latest and keeps your secrets. Add LIBRIS_TAG=0.29.0 to stay on this exact release. An untouched Compose file is updated; a customised one is never overwritten: the installer writes docker-compose.yml.libris-new next to it and stops so you can merge it. 0.13 includes three database migrations; WebDAV stays off until WEBDAV_HOSTS is set, LDAP is turned on from Settings, and an existing MAX_UPLOAD_MB=60 stays as it is. 0.14.0 migrates the usage counters so each total follows its signed licence ID. 0.15.0 has no database migration; it adds a logs volume for the journal and renames the containers libris-api, libris-worker and so on, so update any script that used the old names. Providers that members add for themselves can now reach public addresses only: list a LAN range, for a local vLLM host for example, in PROVIDER_MEMBER_PRIVATE_NETWORKS. From 0.15.2 the installer adds COMPOSE_PROFILES=codex to .env, so a plain docker compose pull also updates the Codex bridge. 0.16.0 includes two database migrations, for quota cycles and for automatic WebDAV sending. A book that does not fit in what is left of the quota is now refused before it is queued, with the code licence_quota_insufficient: scripts that start books through the API or MCP should handle it. With a Trial or Personal licence, the automation API, MCP and sharing stop working; tokens and links are kept. Automatic WebDAV sending is off for every existing connection. Run the installer again rather than docker compose pull alone: an installation left on an older Compose file is now reported as outdated. 0.16.1 includes one database migration: it marks stale the passage scores that counted failed model calls or arbitrations, and they are recomputed the next time the book’s quality or editor is opened. The interface, the e-mails and the OpenAPI title now say Libris Translate; new authenticator codes and security keys are enrolled under that name, those already registered keep working, an explicit WEBAUTHN_RP_NAME is kept, and technical names (libris, LIBRIS_*, images, librisctl) do not change. Its licence refusal messages follow the codes of licence server 0.8.1 or later. 0.16.2 has no database migration. 0.16.3 renames the Compose project libris: the next update stops Libris while it copies the epub-translator_* volumes into libris_* twins, so keep room for one more copy of the data; the old volumes are never removed. 0.16.4 has no database migration; the images now come from registry.libris-translate.com/libris/libris, which the installer writes into .env by itself, and your registry credentials do not change. 0.17.0 and 0.17.1 have no database migration and count nothing again: the cycle counter goes on from the number it had. From 0.17.0 a book is counted as soon as it is added, all of its words at once; books already there keep being counted as they are translated. A book that does not fit in what the cycle has left plus the margin is refused whole, with 402 licence_quota_insufficient, or 402 allowance_insufficient for an account’s own quota, and nothing is imported: scripts that add books through the API or MCP should handle it. From 0.17.1 librisctl also recognises an installation made by the installer, in /opt/libris. 0.18.0, 0.18.1 and 0.19.0 have no database migration. From 0.18.0 the French the machine writes gets the new typography at its next rewrite; existing translations stay as stored. From 0.18.1 AUTOPILOT_ESCALATE_AFTER and AUTOPILOT_ESCALATE_WINDOW no longer exist and are ignored if left in .env. From 0.19.0 a question asked about a passage needs a valid licence, like every other model call; without one it answers 402 with the code licence. 0.19.1 has no database migration. It explains when an automatic trial has already been used on this machine or database; the licence server must be at 0.12.0 to send that refusal. 0.20.0 has no database migration. The application package in the Docker image contains .pyc files instead of .py sources; scripts that read those source files need a source checkout. Error tracebacks still show file names and line numbers, but no code lines. 0.21.0 has no database migration and works with licence server 0.12.0. The quality judge stays off until you choose one in Settings › Autopilot (or QUALITY_JUDGE_PROVIDER); with a judge, each passage costs one more model call on its provider, more for the passages it corrects. The built-in prompts move to file-v9+rules-v6: a prompt an administrator saved in Settings › Prompts keeps its own text. Books already imported keep the lines of their source site. 0.22.0 has no database migration and works with licence server 0.12.0. Its built-in prompts move to file-v11+rules-v6. Its prompts are larger on large windows: the optional context now takes three quarters of what the model's window leaves, up to a Context budget of 32 000 by default (a saved 12 000, the former default, reads as 32 000), about 5 000 more real tokens per call on a 64k window; save 11 999 or less to keep the former size. Books already analysed are dated as they stand, nothing to run again. 0.23.0 includes three database migrations, the longest counting the words of every existing passage (a few seconds for a large library), and works with licence server 0.12. Its built-in prompts move to file-v13+rules-v6: a prompt an administrator saved in Settings › Prompts keeps its own text. 0.24.0 includes one database migration and works with licence server 0.12. Run the installer again rather than docker compose pull alone: it installs next to Compose the backup and restore tools the image now carries. Every translated or rewritten passage costs one more model call on its provider, for the independent semantic check; when that check cannot run, the new version is refused and a review point opens. Its built-in prompts move to file-v16+rules-v7: a prompt an administrator saved in Settings › Prompts keeps its own text. 0.25.0 and 0.26.0 have no database migration. A Personal licence may now show the date through which it includes new versions; check the licence page before installing a release. 0.27.0 includes two database migrations that run on their own at start-up and works with licence server 0.13. Quality alerts raised before the upgrade stay open and listed for review, but cost no points until a later check finds them again. Its built-in prompts move to file-v30: a prompt an administrator saved in Settings › Prompts keeps its own text. 0.27.1 has no database migration; its built-in rules move to rules-v13. 0.27.2 has no database migration either. Its Codex bridge image moves to Codex CLI 0.160.0, which the installer pulls with the application, so that a ChatGPT account is offered the models OpenAI added since. The scheduled backup now pauses the API and the worker while it runs: translations stop for the length of the backup and resume by themselves, and LIBRIS_BACKUP_FREEZE=false keeps the former behaviour. 0.28.0 includes one database migration: it recomputes the translation-memory key for passages containing verse; all other keys stay the same. The quality page can show resolved alerts again, unmet locked glossary terms are reported as residual points instead of blocking a book, and the configured quality judge reads semantic-fidelity checks. 0.29.0 has no database migration and works with licence server 0.13.0. Its application and Codex bridge images include Poppler and Tesseract for local PDF extraction and OCR; update both images with the installer. PDF passwords are used for extraction only and are never retained. Image rollback is not schema rollback: to go back, restore the backup taken before the upgrade.

Then translate a first book

  1. Settings › LLM providers: add your model's endpoint, model name and key.
  2. Add content in the library: pick an EPUB or chapter files and choose a series.
  3. Choose the provider and the languages, then Import and run the whole pipeline.
  4. Follow the progress on the book page and download the result when it is ready.

Only import books you have the right to translate. Quality, supported languages, speed and cost depend on the model you use.

Version 0.29.0

What recent versions added

The latest versions. The changelog lists every change, version by version. Read the full changelog

The speed gain of parallel analysis was measured on a reproducible synthetic benchmark. Its literary quality on a real bilingual book remains open, so Libris Translate does not claim that parallel mode translates as well as sequential mode.

Found a vulnerability? Report it privately to licences@libris-translate.com with [SECURITY] in the subject. Leave out secrets, session cookies and book text.

Changes in the latest releases

Import PDFs, including scans

Version 0.29.0 imports text PDFs, scans and mixed documents, including rotated or password-protected files. Local OCR runs automatically; imports and their archives never retain the password or repeat OCR on restore.

Autopilot recovers from rejected corrections

In 0.29.0, a correction refused by the checks gets one informed retry in the same round. A second refusal stops retries, and the point follows the release's end-of-rounds rules.

Autopilot handles malformed decisions safely

In 0.29.0, one malformed arbitration decision no longer discards the other valid decisions. Each is settled on its own, while missing or unknown decisions still refuse the answer.

PDF extraction stays local

PDF text extraction and OCR run on your Libris server with Poppler and Tesseract. Page layout and images are not reproduced, and password-protected PDFs require a password for extraction only.

Support

Write to support@libris-translate.com from the address your licence was sent to, with your Libris version and what you see. We answer on business days, Paris time. The times below are the first-reply times we aim for: indicative, not contractual, and no compensation is due when a reply takes longer.

PlanFirst reply (indicative)
Trial Best effort
Personal 3 business days
Studio 2 business days
Pro 1 business day

Security report: acknowledged within 2 business days, whatever the plan. Write privately to licences@libris-translate.com with [SECURITY] in the subject.

What support covers

  • The latest published version of Libris and the one before it, installed with the official installer.
  • Installation, activation and the licence, updates, errors and anything the software does that the documentation says it should not.

What it does not cover

  • The language model or provider you chose: its answers, availability, limits and prices.
  • A Compose file you have modified, and versions older than the previous one.
  • Your server, network, operating system and Docker installation.
  • Backups of your installation, which are yours to make: we back up the licence server only.

No availability figure is promised for the licence server. Your installation holds a signed certificate valid for 72 hours and renews it every hour, so an outage shorter than that does not stop your work.

Support in the terms

Questions

How much does it cost?

There are two separate costs. Libris Translate itself: 50 € once for the personal plan, for non-commercial use, 29 € a month or 290 € a year for Studio, or a quote for Pro, after a 30-day trial. The model, which you bring and pay directly: a hosted API bills per token, a server of your own costs its hardware and power. Before a paid operation, Libris shows the estimated tokens and cost. Our prices include VAT. You can give each book and each API token a budget; at 90 % of it, Libris switches to a cheaper provider from the fallback chain or pauses the job. A passage identical to one you already translated skips the translation call, though it still counts toward the word quota. What a real book cost, model by model

Can I try it before buying?

Yes, for 30 days; the form on this page sends a confirmation link by email. Follow it within 24 hours to receive your key. The trial has every format, every provider and the full autopilot, with 100 000 words a month, one account and two machines. The automation API, the MCP server and book sharing (members and review links) come with Studio and Pro. One trial per address, for life, and no payment card. When you buy, the trial becomes your licence on the same key: the machine stays activated, the quota starts a new monthly cycle on the day you buy, and there is nothing to type again. What the free trial includes

What happens if my licence expires or I stop paying?

From the due date, Libris Translate stops starting new translations. Everything already produced stays readable, exportable and deletable, and you can still correct it by hand: the library, the translations, the glossaries and the reports. You can leave with all of it. Access to the image registry, and so to new versions, ends 7 days after the due date. End of a subscription, in the terms

Can I move Libris Translate to another machine?

Yes, yourself, from the licence screen. A personal licence runs on three machines at once, Studio on five and Pro on ten. Releasing a machine frees its slot at once. If you activate one more machine than your plan allows, the oldest one is released automatically, provided it has held its place for at least seven days; that rule is what stops a single key from being passed around.

What do you see of my books?

Nothing of their content. Once an hour, and when a job ends, your installation sends our licence server its key, the Libris Translate version, the words counted in the current quota cycle, the number of active accounts, an installation ID and a hashed fingerprint of its machine, signed by the installation. No title, no text, no character name. The passages being translated go to the model provider you configured, and nowhere else.

Which models can I use?

Any server that speaks the OpenAI-compatible Chat Completions protocol, local or hosted, plus the OpenAI and Anthropic APIs. An optional bridge lets you translate with Codex through a ChatGPT account. Each provider has its own concurrency limit and prices, and can act as a fallback for another. Which model to choose, and how to get a key · Libris and the alternatives

Which languages?

Whichever your model handles well, including Tamil: quality and language coverage depend on the model. Books translated into Arabic, Hebrew, Persian, Urdu, Pashto, Yiddish or another right-to-left language are exported right to left. The Libris Translate interface is available in English and French.

Do my books leave my server?

Your books, translations and history are stored on your server. Analysis, translation and review send passages and context to your configured model providers; a hosted model receives that text. Optional memory, web search, mail and delivery send the data their enabled features need. You can use a local model and leave those integrations off. API keys are stored encrypted.

Do I need a GPU?

Not for Libris Translate itself. A Linux (amd64) machine with Docker and 2 GB of free memory is enough when the model runs elsewhere, on a hosted API or on your own inference server.

Which formats are supported?

In: EPUB 2 and 3, chapters as TXT, Markdown, HTML or DOCX (a TXT, Markdown or DOCX file can be split at its chapter headings), text PDFs, scans and mixed PDFs with local OCR, and JSON through the API. Out: the translated EPUB with its original layout for an EPUB source, or a new EPUB built from the chapters for any other source, checked with EPUBCheck; a bilingual EPUB for proofreading; a Word document with your corrections as tracked changes and the model's doubts as comments; text, Markdown, or a ZIP of chapters. Fixed-layout EPUBs are refused at import, and DRM-protected EPUBs are not supported.

Can I follow a webnovel that is still being published?

Yes. Send each new batch of chapters to the same series: they are placed by number, duplicates are skipped, and only new or replaced chapters are translated, with the earlier ones as context and no new model call for those. Chapters published in parts, prologues, interludes and side stories keep their place in reading order.

Can a coding agent drive Libris Translate?

Yes. Libris Translate answers the Model Context Protocol at POST /mcp, so opencode, Claude Code, Codex or any other MCP client can read the library, set up a book, edit the glossary, start and steer jobs and read the translation back, without hand-written curl calls. It uses the same account, the same audit log and the same rules as the interface. Give the token only the scopes the task needs: a token that can read a library cannot empty it, and every deletion must be confirmed explicitly. Studio and Pro include it.

Can several people share one installation?

Yes, up to the number of accounts in your plan: one for Personal, three for Studio, ten for Pro unless your quote says otherwise. Each person has their own account and, with Studio and Pro, their own API tokens. Waiting jobs start by priority, then take turns between accounts, and administrators can set quotas per account and per token. Studio and Pro also share books between accounts, and a proofreader needs no account at all: a review link opens one book, read-only, uses no account seat and can be revoked at any time.

Is there a hosted version?

Not yet. Libris Translate is software you install and run on your own machine, and we never hold your books. A hosted option is planned; this page will announce it when it opens.

What if the publisher stops?

Every valid licence then receives an offline certificate that never expires and has no word limit, so Libris Translate keeps working without our licence server, and the image registry stays open for 6 months so you can install the version that accepts it. Read the commitment

What support do I get?

Email support on business days, Paris time, with indicative first-reply times that depend on your plan. They are not contractual. Support by plan

What is the licence?

Libris Translate is proprietary software. Your key lets you install and use it on the number of machines and within the limits of your plan. What you translate is yours: Libris places no restriction on the books you import, their translations or the files it produces. Respecting the rights to the original works is your responsibility. A valid licence gives access to every version published until the publisher ceases trading. The details are in the terms of sale and licence.

Your next series, translated on your own server.

Try it free for 30 days, or install it with Docker in about twenty minutes.