Skip to content

Translate it.Keep it.

For software teams, localization owners and technical writers who move product content, documentation and resource files between languages. 
Windows open-source desktop application.
Not a browser translator and not a cloud service at this time.

Windows 10 / 11 AI runs on your PC

Build v1.0.7 / 8afffb9 / 2026-09-04Downloads (78)

TRUSTED PARTNERS

01Workspace

Five places your work passes through.

Work moves through one window, start to finish. Nothing leaves the machine on the way.

  • Recent chatsEvery chat and entry, kept in one local database.
  • The composerType a line or drop a file.
  • Advanced panelOne control sets the model and the token budget.
  • Project memoryAdds your glossary and passages from the indexed project.
  • Your dataWhat is stored here, each row with its size.
  • Software teams
  • Localization owners
  • Technical writers
The BetterTranslator window with no translations yet. A sidebar of recent chats, an empty message area, and the composer with From English and To Czech.
02Formats

Six formats in, two ways back.

Word and PDF cannot be written back, so both come back as text.

Same file back

Markdown.md.md
JSON.json.json
Text.txt.txt
Subtitles.srt.srt

Back as text

Word.docx.txt
PDF.pdf.txt
Excel.xlsxnot supported
See a sample
Markdownsame format back, fences and tables kept
ENguide.md
## Getting started

Run `bt --help` to list every command.

- Install the app
- Choose a folder
CSguide.cs.md
## Začínáme

Spusťte `bt --help` a zobrazí se všechny příkazy.

- Nainstalujte aplikaci
- Vyberte složku
03How it works
EN Čeština On your machine

Translate {count} files locally, keep every <a>tag</a> and every BetterTranslator name unchanged in the output.

Přelož 4 soubory lokálně, zachovej každý tag a každý název BetterTranslator beze změny ve výstupu.

Accepted 5 Retried 1 Source kept
04Savings

The whole file, zero API tokens.

Read the benchmark and re-run it on your own file.
Fine print

Benchmark run 20260812T160551Z-92885b, captured 2026-08-12. Source file: an en.json resource file of 4,666 keys and 627,920 bytes, sha256 a8fdce5a, translated English to Czech. Agent: Claude Code CLI 2.1.228 on claude-opus-5, fresh session per run, three repetitions per cell with the first excluded as cache warm-up. Local engine: TranslateGemma (translategemma-4b-it-Q4_K_M.gguf, Q4_K_M) through BetterRuntime on Windows 11.

Measured slices hold 10, 50, 200 and 500 keys. Whole-file figures are projections: the 200 key slice repeated as 24 chunks, arithmetic over measured runs rather than a measured run.

Cost figures are approximations reported by the Claude Code CLI and are not billing figures.

Above about 87 keys the delegated path costs less money than the agent alone; below that, the agent alone is cheaper.

Two results run against the product. The local engine left 2.3 percent of values in English on the slices it was measured on, and no adequacy or fluency judgement was made on the values it did change. At 500 keys the delegated path produced no file in any repetition, because the agent stopped waiting before the local job finished and the job did not outlive the session.

05Terminology

One term, the rendering you accepted.

TermAcceptedRejectedReason on record
buildbuildstavbaa noun in your product, not construction
commitcommitzávazekkept in English by the table
issueúkolproblémthe tracker’s word, not a complaint
06Effort

One control, until you want the panel.

Simple

SimpleEuroLLM, 512 tokens at most
ThinkingTranslateGemma, 1536 tokens at most

Advanced

ModelTranslateGemma 4b-it-Q4_K_MEuroLLM 9B-Q4_K_M
Temperature0
User promptKeep interface labels short enough for a 180 pixel button.
Per message6 tokens · 3.2 s · 1.9 tok/sCoverage100%1 line · 2 units
07Automation

Command Line Integration / MCP

  • translate_text
  • translate_file
  • translate_batch
  • job_status
  • job_cancel
  • list_languages
  • list_models
  • select_model
  • show_in_gui
  • get_entry
Agent session
> bt mcp
  stdio server ready. ten tools, no port opened> Translate the English docs folder into Czech.

  translate_batch
    path: docs\en
    to: cs

  job 1 queued
  job_status 1  running
  job_status 1  finished

  docs\cs> bt translate --to cs --file docs\onboarding.md
  docs\onboarding.cs.md> bt translate --to cs --batch docs\en --out docs\cs
  exit 0
08Download

Download it and translate a file locally.

First run installs a model. After that the model loads from your disk into a process on your machine, and the only outbound request is behind you.