What a model provider is
A model provider is a saved credential for one LLM vendor, with the scopes that can use it. LangWatch uses it whenever a feature calls a model on your behalf: online evaluators, Workflows, the prompt playground, Agent Testing, topic clustering, embeddings, trace AI search, Langy and the LangWatch AI Gateway. Manage them at Settings > Model Providers. The table lists each configuration with its Scope chips, broadest scope first. The row menu has Edit Provider, Test Connection and Delete Provider. On a self-hosted deployment, a vendor whose key is set in the server environment shows as a row with a System chip and no menu.
Supported providers
Add Model Provider opens a menu with one entry per vendor:
You can save several configurations of the same vendor, for example one OpenAI account per region, and tell them apart by Name.
Add a provider
1
Pick the vendor
Click Add Model Provider and choose the vendor. The drawer opens.
2
Name it
Name is required, up to 128 characters. It is how the configuration appears in model pickers and in the LangWatch AI Gateway.
3
Pick the scope
Scope is a multi-select of the organization, teams and projects. It starts on the widest scope you can manage. The helper text states who can use the configuration, for example “Only this project can use this provider.” You need
organization:manage, team:manage or project:manage for each scope you pick.4
Enter the credentials
Fill in the vendor’s fields. Below them: Extra Headers sent with every request, the routing handle line (“Requests reach this provider as
openai/<model>”) for the LangWatch AI Gateway, Custom Models, and Advanced (Gateway) for rate limits and fallback priority.5
Save and test
Save, then open the row menu and click Test Connection. The verdict appears under the provider name in its row: Connection works, the vendor’s refusal message, or a note that the provider cannot be tested automatically and is checked on first use. The test runs on the server with the saved key; the browser never receives it.
Custom models
Every LLM provider has a Custom Models section. Click Add, choose a chat model or an embeddings model, and enter the model id. The entry appears in the model pickers next to the vendor’s known models, which See all models lists. Use it for fine-tuned models, new releases and the models behind a Custom (OpenAI-compatible) endpoint.Who can see and use a configuration
A project sees every configuration scoped to the project itself, to its team or to its organization. Each one is a separate entry in the model pickers under its own name.
The All you can see filter above the table narrows the list and the default models to one scope. Every save and delete is written to the Audit Log as
modelProvider.update or modelProvider.delete.
Default models
The Default Models section on the same page sets the model each feature uses unless the user picks one in the feature itself. A config sets a model per role (Default, Fast, Langy, Embeddings) and per feature, for example the model a new prompt or evaluator starts with, the Langy chat model, trace AI search, workflow commit messages, and the topic clustering model and embeddings. Click Add config, pick the scopes and the models. A project config beats a team config, which beats an organization config. A feature that no config names inherits from the wider scope.