Localization engines

Create shared translation configuration, select it in your callers, and manage the resources and workflows that depend on it.

A localization engine gives recurring translation requests a shared configuration. It selects models and links to the terminology, instructions, and voice that the work needs. An application can send its content without repeating every product convention in each request.

In the Studio 2.5 navigation, open Studio 2.5 → Localization engines. Otherwise, open Infrastructure. Existing engine links still work.

The localization engine belongs to an organization. Its attached glossaries, rulesets, and brand voices also belong to that organization and can be shared with other localization engines. That separation lets related workflows use common terminology while keeping different model choices or pipeline policies.

Create the first configuration #

  1. Select the intended organization, open Infrastructure, and create a localization engine. Give it a name that identifies the work, such as Product documentation.
  2. On Models, review the recommended primary model and fallback. Their initial catch-all configuration covers all locale pairs. Keep it for a first trial or configure the model coverage the work requires.
  3. If Context appears, select existing glossaries, rulesets, and one brand voice, or leave the selections empty. If Website appears, enter a site to generate missing guidance, or select Skip to continue without it. Review the selected attachments and generation plan, then create the localization engine.
  4. Inspect the saved attachments. Creation and attachment are separate writes, so the localization engine can exist even if an intended attachment failed.
  5. Copy its ID from the page header into your caller's supported configuration. Translate a small representative sample and inspect the output and request evidence.

A new localization engine created in the web app starts with a translation-only pipeline. Optional rewriting and human review are disabled, as are glossary/instruction scoring and automatic suggestion generation. Review the saved settings before enabling additional behavior. Existing localization engines may already have different policies.

Configure Human review in the engine's Pipeline tab. Its Decision pools cards choose who reviews each language and directory; Decision pools covers creating pools and setting those routes.

The API quickstart completes the first request using an explicit ID. The CLI quickstart follows a file-based workflow. The CLI manages file submission and retrieval through the same document-job lifecycle; use the configuration supported by your caller.

Select the intended localization engine #

Every API translation job requires an engineId. The CLI saves that selection with lingo link; integrations save it in their own configuration. Confirm the owning organization and intended guidance before submitting work. A display name helps people recognize the engine; callers use its identifier.

For each request, the localization engine resolves model candidates, matching instructions, one voice variation, and relevant glossary terms. These resources do not all follow the same matching rule. Locale resolution explains how they accumulate, compete, or are filtered before use.

Develop shared guidance #

Use Translation context to decide whether a requirement belongs in terminology, instructions, or voice. Before editing a shared resource, inspect the localization engines that use it. A change appropriate for one audience may need a separate resource rather than an edit affecting every consumer.

You can also configure resources through the account MCP tools. Ask an assistant to show proposed terms, instructions, and locale-specific voice before applying them. Generate guidance from existing content describes asynchronous provisioning from URLs or text samples. Both approaches need the same saved-record and output inspection as manual configuration.

Keep configuration changes with the request evidence when explaining work that spans an edit. A saved attachment identifies a shared resource; it does not by itself preserve every version of that resource's text. Record the relevant guidance before a comparison instead of assuming today's settings reproduce an earlier request.

Disable or delete a localization engine #

Disable a localization engine to stop new execution reversibly. First identify the callers using its ID and decide whether to pause or redirect them. Disabling does not cancel all in-flight jobs or already queued suggestions, and it does not undo translations or guidance changes already applied.

Delete only after removing or replacing dependent caller references. Requests naming the deleted ID can no longer use it. Deletion removes the localization engine and its attachments, but the organization's glossaries, rulesets, and brand voices survive as shared resources.

Before either action, inspect pending work in Inbox and run evidence. Resolve or deliberately account for that work separately from stopping new requests.