How do you QA a language nobody on your team speaks?

If you ship in a language nobody on your team reads, you have a class of bug you cannot see.

Most of us handle this the same way. Dump the UI strings, the docs, the onboarding emails, the pricing page into whatever model is already in the stack. Get fluent output back. Ship it. Move on to something you can actually verify.

The failure mode is not gibberish. Gibberish would be great, because you would catch it. The failure mode is fluent and wrong:

Formality slips. Your Spanish onboarding uses tú on a page selling to enterprise procurement teams, and your German signup form uses du where every competitor uses Sie.
A CTA turns into the wrong speech act. "Get started" becomes an order in one language and a vague suggestion in another. Same button, different conversion rate, no way to tell why.
False friends land in your billing copy. Spanish actualmente is not "actually." Your refund policy now says something you did not write.
Idioms get translated literally. "Ship fast" reads as a freight instruction.

None of that shows up in a diff. It shows up as a market where your funnel converts at a third of your English rate and you assume the market is just bad.

The part that's useful whether or not you use our thing

Run the same string through several models independently and look at where they disagree.

Where they all land on the same output, you are probably fine. Where they scatter, one of two things is true: your source string is ambiguous, or that phrase is genuinely hard in that language. Either way you now have a ranked QA list instead of 4,000 strings and a vague sense of dread.

It catches source-side problems too. A string that five models render five different ways is usually a string a human reader would also find ambiguous, which means your English copy needs work before anything gets translated.

You can build a crude version of this in an afternoon. Three API keys, run each string through all three, flag anything where the outputs diverge beyond a similarity threshold. No product required.

What we built

We took that further than three keys. MachineTranslation.com runs text through 22 AI translation models at once and returns the consensus rather than one model's guess. We put it behind an MCP server so your agent can call it directly instead of you round-tripping through a web UI.

Two tools:

smart_translate (text, source language, target language)
list_languages (333 languages)

Server URL: https://www.machinetranslation.com/mcp

Add it to Claude, Cursor, or any MCP-compatible client, authenticate with your account, and your agent can translate mid-task. You ask it to localize a component and it returns strings that 22 models agreed on, without you switching windows.

Straight answer on cost, because I'd want it. The MCP requires a paid plan. There is no free tier on it right now. Docs are at developer.machinetranslation.com/mcp if you want to read the tool schemas before deciding whether it's worth it.

Full disclosure, I work on MachineTranslation.com. Use the disagreement trick anyway. It costs you nothing and it will find real bugs in your localized copy.

What's the worst mistranslation that made it into your product?