Use this article when you want wxrks's AI steps to run on an engine of your own: an LLM account your company already has, a model you host, or a dedicated post-editing or quality-estimation service. By default, all operations use wxrks's own default models, but any operation can be routed to a different model, which overrides the wxrks model for that operation. This article explains which engines you connect yourself, which ones wxrks connects for you, and which AI operations you can reroute.

Who is this for? Account Admins. Custom AI engines are configured in Model Management, which requires the wxrks FUSION plan or higher.

The translation's Confidence Score is based on Smells, and Smells runs on an LLM. The model reads each source and target segment and reports the quality problems it finds, as findings with a category and a severity. wxrks then turns those findings into the score: a segment with no findings scores 100%, and otherwise the single most severe finding sets the score.

ℹ️ Note: Because an LLM finds the problems and the score follows from them, whichever engine runs Smells determines the score. A different model can report different findings for the same segment, and so give a different score. The Overall Confidence of a work unit is the plain average of its segments' scores. For how to read the score in the Editor, see QA Confidence Score and Smells.

Go to Settings > Account Settings > Model Management and click Add Custom Setting. The providers you can connect there are OpenAI, Azure OpenAI, Anthropic, Groq, OpenRouter, Cerebras, Google Gemini (Studio) and Google Gemini (Vertex AI). The full walkthrough is in Configure Custom AI Providers with Model Management.

Add Custom Setting (1) creates a row for your own engine. The Managed rows at Global scope (2) are wxrks's own default models. They come last in the order of precedence. If no other model is defined for an operation, wxrks uses its own model for it. If an operation is set to a different model, the model you chose overrides the wxrks model for that operation. The wxrks Default row, which has no operations, is the last-resort fallback.

What you enter depends on the provider:

Your own LLM behind an OpenAI-compatible endpoint. This is the simplest way to use an engine of your own, and it needs no work from wxrks. An OpenAI, Groq, OpenRouter or Cerebras row can point its Base URL at any endpoint that speaks the OpenAI chat-completions format, such as a gateway in front of a model you host. The API key is sent as a bearer token. The endpoint must:

Any provider or protocol outside the self-service list is connected by the wxrks team in the back-end.

To request a connection, contact your Account Manager. Say that you want to connect a custom AI engine, and send the items listed below.

Depending on the complexity of the implementation, wxrks will submit a quote for your approval before the work starts. Once the scope is agreed, implementation takes about 1 week.

💡 Tip: Check first whether you need a back-end connection at all. If your engine accepts the OpenAI chat-completions format with a bearer API key, connect it yourself, as described above.

What wxrks needs is different from a custom MT engine. wxrks talks to every LLM in the same way: it sends its own prompts in chat format and reads the model's reply. So wxrks needs to know how your service behaves in that exchange, not only where it is:

What wxrks does. The wxrks team adds your provider to the list of supported providers, builds the connection to it, adds the credential checks and usage tracking, and then makes it available in Model Management.

ℹ️ Note: wxrks connects to an engine with an API key over a public HTTPS endpoint. wxrks handles every AI operation as a chat exchange, so a service that is not chat-based, for example one that takes a segment in and returns a post-edited segment, needs a new adapter built by wxrks. If your engine needs a different authentication method or a private network connection, say so when you make the request. wxrks evaluates these case by case.

Each Model Management row has a scope and a set of operations:

For each call, wxrks uses the most specific row: an Org Unit row for the exact operation, then the Org Unit default row, then an Account row for the exact operation, then the Account default row, then the wxrks default models.

These are the values in the Operations field. Any of them can be routed to your engine, and a row of yours overrides the wxrks model for that operation.

What each quality operation covers:

ℹ️ Note: The Confidence Score comes from Smells. If you route QA Smells to your own engine, your engine scores the translations. If you route CS Translation but not QA Smells, your engine writes the translation and the wxrks model scores it.

There's no switch that turns off the wxrks models for an account. To get as close as possible:

⚠️ Warning: A row that can't be used — because it's disabled, or its configuration is incomplete — is skipped without any message, and the call falls through to the next rule, eventually the wxrks default models. Keep your rows Enabled, and check the Usage column to confirm your row is the one taking traffic.

ℹ️ Note: If your engine errors or times out during a call, wxrks does not retry the call on its own models: the action fails and the error is shown. For CS Translation, a row can list fallback models that are tried on the same provider when the primary model is rate-limited or unavailable.

Providers you connect yourself. The credentials of an AI provider are tested when you save the row in Model Management:

A new provider that wxrks connects for you. The wxrks QA team tests the connection in the DevOps environment before it goes live, using the test key or test environment you sent (see the table above).

To change an engine that wxrks connected for you, such as its endpoint, authentication, models or credentials, contact wxrks Support.