Context tools for AI translation: glossary, style guide, memory & screenshots

Automation already handles the mechanical side of localization well: keys get reported via saveMissing, the CLI instruments your code, CI publishes on deploy, and Automatic Translation fills the gaps. But when an AI coding agent translates or reviews a string itself, right there in your editor, it has been missing the context a human translator takes for granted: which term is the approved one, what tone the brand uses, how this string was translated last time, and where it even appears in the UI.

Four new read-only tools on the Locize MCP server close that gap.

Lint terminology in a pull request, no translation needed

Start with the one that earns its keep even if you never let an agent translate a word: get_glossary as a terminology linter.

Your Locize glossary already knows which terms are approved and which are forbidden in each language. Point an agent at a pull request that adds or changes English source strings and ask it to check them:

"Lint this PR's new English strings against the project glossary and flag any forbidden terms."

The agent calls get_glossary with the source language, gets back the forbidden set, and flags drift before it ships, suggesting the preferred term in its place. Terminology consistency, enforced at review time, with zero translation involved.

// get_glossary(projectId, languages: "en")
{
  "sourceLanguage": "en",
  "terms": [
    {
      "id": "…",
      "title": "login",
      "languages": {
        "en": { "preferred": "log in", "allowed": ["sign in"], "forbidden": ["log on"] }
      }
    }
  ]
}

Translate with the glossary, style guide, and memory

When you do let the agent translate or review, three tools give it the rest of the context:

• get_glossary with languages: "en,de" returns the preferred target term per concept, so the agent emits the sanctioned terminology and avoids forbidden variants.
• get_styleguide returns the project's tone, formality, target audience, and usage rules per language, so the translation matches your brand voice instead of a generic default.
• search_translation_memory finds previously translated segments similar to the new string (exact and fuzzy, across the project and any TMX-linked projects), so the agent reuses an existing translation instead of inventing a new one.

"Translate the new checkout strings into German using the project's glossary, style guide, and translation memory."

The result reads like the rest of your product, because it was produced with the same inputs your translators use, the same inputs Locize's own Automatic Translation already feeds to the AI/MT providers.

See the string before translating it

Some strings are impossible to translate well without seeing them. A bare "Open" could be a button, a status, or a menu item, and the right translation, and the available space, depend on which.

get_screenshot_context maps a translation key to the screenshot(s) where it appears, with the highlighted region and any OCR-extracted text. The agent (or you) can fetch the image to disambiguate meaning and check for overflow before committing a translation.

"Show me the screenshot where the checkout.title key appears."

Setup

Nothing new to install. If you've already connected the Locize MCP server, the four context tools are available now. They're read-only and need only the read PAT scope. If you haven't, setup takes about 60 seconds via OAuth or a Personal Access Token.

The full tool reference is in the MCP documentation.

To make your agent do this automatically, drop our ready-made conventions (CLAUDE.md, AGENTS.md, .cursorrules) into your repo — they're in the public locize-agents repository.

Why this matters

The mechanical parts of localization are increasingly automated. What's left, and what actually determines quality, is context: terminology, voice, prior decisions, and visual placement. By exposing that context over MCP, Locize lets your AI agent translate, review, and lint to your project's standards, in the tools you already work in.

Try it on your next PR: ask your assistant to lint the new strings against your glossary. It's the fastest way to feel the difference.