i18n Agent
A multi-model AI translation platform that turns localization files into finished translations in minutes, across 100+ languages, without breaking the code around the text. We designed it, shipped it, and run it in production every day.
What it is
i18n Agent is a professional AI translation platform for teams that ship software in more than one language. You upload your localization files, pick the target languages, and get back translations that read naturally and keep every placeholder, tag, and structural key exactly where it belongs. It supports 100+ languages and handles the file formats developers actually work in: JSON, YAML, CSV, XML, Markdown, Properties, and more.
The people who use it are developers and localization owners at companies from startup to enterprise scale. They reach for it because a traditional agency quote runs into the thousands per project and takes weeks, while the same work through i18n Agent completes in minutes at a fraction of the cost. It also plugs into the tools those developers already live in, with integrations for Claude Code, Cursor, and VS Code.
See it live at i18nagent.aiWhy we built it
Software localization is a genuinely hard problem dressed up as a simple one. The naive version is "send the strings to a translator and paste the results back." In practice, a single mistranslated interpolation token or a stray closing tag can break a build or ship a broken screen to users in another market. Traditional agencies are slow and expensive precisely because a human has to protect that structure by hand, one file at a time.
We built i18n Agent because we kept hitting this ourselves and could not find a tool that treated translation as an engineering problem rather than a text problem. General-purpose chatbots translate the words but happily rewrite your JSON, drop your ICU plural forms, or "helpfully" localize a variable name. We wanted a system that understands the difference between the text a user reads and the structure a program depends on, and never confuses the two.
The goal was straightforward and demanding: translations good enough to ship without a second pass, fast enough to run inside a release cycle, and cheap enough that translating into another ten languages is a decision you make casually rather than budget for over a quarter.
How we built it
The core of i18n Agent is a multi-model orchestration layer. Rather than betting the whole product on one large language model, we route work across GPT-5, Claude, and specialized models, choosing the right engine for each stage of the job. That routing is where a lot of the quality and the cost efficiency come from: the strongest model is expensive and you do not need it for every token, so the pipeline spends it where it changes the outcome and uses lighter models everywhere else.
Structure-aware parsing
Files are parsed into a representation that separates translatable text from the keys, placeholders, and markup around it. The models only ever see and rewrite the parts that are meant to change.
Multi-model routing
A routing layer sends each unit of work to GPT-5, Claude, or a specialized model based on the language pair and the difficulty of the content, balancing quality against cost per word.
Context-aware translation
Translations are produced with surrounding context so the same term stays consistent across a whole project, instead of drifting file by file the way isolated string translation does.
Format fidelity
The output is reassembled back into the original format so JSON stays valid JSON and XML stays valid XML, with every placeholder and tag preserved.
The hard parts were not the individual translations, they were everything around them. Keeping terminology consistent across hundreds of files without a human editor. Handling plural rules and gendered forms that differ language by language. Deciding, per request, which model to trust so that quality holds without the cost running away. And doing all of that reliably enough to expose it as a service other developers build on top of, including directly inside their editors through the Claude Code, Cursor, and VS Code integrations.
Underneath, it is a TypeScript and Next.js application running on AWS, with the translation pipeline built as a set of services rather than one monolith, so each stage can be scaled and improved independently. It is built to enterprise expectations on the boring but essential details too, with encrypted transport and GDPR-compliant handling of customer content.
What this proves for your project
i18n Agent is not a client case study we are borrowing credit for. It is our own product, which means the hard engineering decisions in it are ones we made and now live with. Here is what that experience transfers to your build.
Multi-model LLM orchestration that respects a budget
Most teams either overpay by sending everything to the biggest model or underdeliver by cheaping out on all of it. We have built and operate a routing layer that spends the expensive model only where it matters. If your workflow involves AI, that is the difference between a demo and a system with defensible unit economics.
AI pipelines that hold up in production
Structure-aware parsing, format fidelity, and consistency across a large batch are the unglamorous problems that decide whether an AI feature is trustworthy. We solved them for our own paying users, not in a prototype.
Cost-efficient AI that a business can actually run
Taking a process that cost thousands and weeks down to minutes and a fraction of the price is exactly the kind of outcome we look for in a client engagement. We did it for translation. The method generalizes.
If you want that applied to your own work, start with our AI workflow automation work, see how we build internal tools, or check the honest pricing before you talk to anyone.
Want something like this built for your team?
The senior engineer who built i18n Agent is the same person who would build for you. Start with a fixed-price audit, or just book a call and tell us what you are trying to solve.