Introduction

What is translate-kit?

In 2026, manually writing i18n JSON files key by key is obsolete. AI handles translation now, but ad-hoc prompts create inconsistent results across files. translate-kit is a single tool that handles the entire translation pipeline at build time, powered by AI, with zero runtime cost.

It extracts translatable strings from your JSX/TSX, generates semantic keys with AI, replaces hardcoded text with i18n calls, and translates everything to your target locales — using your own AI models via Vercel AI SDK. Compatible with next-intl.

What Makes It Different

• Build-time, not runtime — translation happens before you ship. No client-side SDK, no loading spinners, no runtime overhead. The output is static JSON files that next-intl serves directly.
• Zero dependency footprint — translate-kit is not a runtime dependency. It generates standard next-intl code and JSON files. If you remove translate-kit tomorrow, your app keeps working exactly the same.
• AI-native, provider-agnostic — works with any AI model from any provider (OpenAI, Anthropic, Google, Mistral, Groq). The AI receives full context: project description, glossary, tone, namespace structure.
• Incremental by default — a lock file tracks SHA-256 hashes. Only changed keys get sent to the AI on re-runs, keeping costs predictable and API calls fast.
• Scanner understands code — parses your AST, distinguishes user-facing text from code artifacts, and filters automatically. No manual string extraction.

How It Works

translate-kit provides a three-step pipeline:

scan → codegen → translate

1. Scan — parse JSX/TSX files, extract translatable strings, generate semantic keys via AI
2. Codegen — replace hardcoded strings with t("key") calls or <T id="key">text</T> components
3. Translate — diff source messages against a lock file, translate only new/modified keys

You can use any step independently or run the full pipeline with translate-kit run. The translate command is the most commonly used on its own — write your source messages manually and let translate-kit handle the rest.

Two Modes

• Keys mode (default) — strings replaced with t("key") calls, source text moves to JSON files. Works with next-intl's useTranslations.
• Inline mode — strings wrapped with <T id="key">text</T>, source text stays visible in your code. See the Inline Mode guide.

Incremental by Default

A .translate-lock.json file tracks source hashes so re-runs only translate what changed. This keeps API calls fast and costs low.