A deterministic, on-device translator for macOS. A UNIX-style command and a drop-in HTTP server for DeepL, LibreTranslate, and Google v2 — all running 100% on-device using Apple's Translation framework.

translate is a tiny macOS CLI built as a UNIX filter: stdin in, stdout out, stderr for errors and optional progress. It uses Apple's on-device Translation framework on macOS Tahoe, auto-detects the source language with NaturalLanguage, reuses a single translation session per run, and batches work wherever possible. With --serve it exposes the same engine over HTTP, byte-compatible with DeepL /v2/*, LibreTranslate /translate /detect /languages, and Google /language/translate/v2/* — so existing client libraries point at it unchanged.

No cloud calls, no API keys, no telemetry, no LLMs, no third-party translation libraries. Apple's Translation framework is a deterministic neural translation engine — same input, same output, every time.

What it is

Requirements

• macOS 26 Tahoe or newer
• Apple silicon Mac
• Swift 6 toolchain (Command Line Tools with the macOS 26.4 SDK)
• Translation language models installed — see docs/install-translation-models.md

Install

git clone https://github.com/Arthur-Ficial/translate.git
cd translate
make install

make install copies the release binary to /usr/local/bin/translate and creates the short alias /usr/local/bin/ueb.

Quick Start: UNIX tool

echo "hallo welt" | translate --to en
translate --to de "hello world" "good night"
pbpaste | translate --to en | pbcopy
translate --to ja --format ndjson < sentences.txt | jq -r .dst
translate --install de-en
translate --detect-only < unknown.txt

Quick Start: HTTP server

translate --serve --port 8989

DeepL

Drop-in for DeepL clients (Python deepl, Node deepl-node):

import deepl
translator = deepl.Translator("any-token", server_url="http://localhost:8989")
print(translator.translate_text("Hallo Welt", target_lang="EN"))

curl -s -X POST http://localhost:8989/v2/translate \
  --data-urlencode "text=Hallo Welt" \
  --data-urlencode "target_lang=EN"
# {"translations":[{"detected_source_language":"DE","text":"Hello World"}]}

LibreTranslate

Drop-in for LibreTranslate clients (libretranslatepy, raw requests):

import requests
r = requests.post("http://localhost:8989/translate",
                  json={"q": "Hallo Welt", "source": "auto", "target": "en"})
print(r.json())
# {"detectedLanguage":{"confidence":97,"language":"de"},"translatedText":"Hello World"}

Google Cloud Translation v2

Drop-in for the v2 REST API. Override the base URL in your client.

curl -s -X POST "http://localhost:8989/language/translate/v2" \
  --data-urlencode "q=Hallo Welt" \
  --data-urlencode "target=en"
# {"data":{"translations":[{"detectedSourceLanguage":"de","translatedText":"Hello World"}]}}

Options

Why on-device?

On-device translation keeps source text local, reduces latency, works offline after models are installed, and needs no API key, account, usage quota, proxy, or billing setup. It is also easier to script safely: the same input, model version, and flags produce byte-stable output without a remote service changing behavior between runs.

Model management

translate --install de-en      # prepare a pair
translate --installed          # list installed pairs
translate --available          # list supported pairs on this OS

For scripts, prevent automatic model preparation:

translate --from de --to en --no-install < input.txt

If the model is missing, the command exits with code 4.

Protected text

translate does not translate fenced code blocks, inline backtick spans, URLs, or email addresses. Those spans pass through unchanged.

Exit codes

Testing

make test                  # unit + Swift HTTP integration tests
make test-real-clients     # exercise the server with real Python clients
                           # (deepl, libretranslatepy, requests, google-cloud-translate)

The unit suite drives every codec and every endpoint via real loopback HTTP. The real-client suite proves drop-in compatibility with three production translation APIs.

See EXAMPLE.md for end-to-end CLI and server scenarios.

Benchmarks

Reference targets on Apple silicon with installed low-latency models:

License

MIT — see LICENSE.