AI models

Every workspace has one AI model configuration. It applies to all agents in the workspace: threads, investigations, automations, autofixes, and background work. Configure it under Settings → AI Models in the console.

Providers

Pick one provider for the workspace:

• Polylane Pure: the default. Polylane manages inference for you. No key, no configuration.
• OpenAI, Anthropic, Google Gemini: managed inference on that provider's model family, with the option to bring your own key.
• Custom (OpenAI-compatible): your own endpoint, your own models. Enterprise plan only.

Any provider other than Polylane Pure requires the Team or Enterprise plan. See Billing and usage.

Bring your own key

With a specific provider selected (OpenAI, Anthropic, or Google Gemini), toggle Bring Your Own Key and paste an API key from your provider account. Model calls then run against your provider account under your own agreement with them, not through Polylane's. Keys are encrypted at rest.

On the Enterprise plan you can also set a Custom Gateway: an OpenAI-compatible base URL that all inference requests route through instead of the provider's default endpoint. The gateway field unlocks after you provide an API key.

Use Test connection before saving. It runs a small test generation through the exact path real requests take, including your gateway if one is set, so a pass means the configuration works end to end.

Switching back to Polylane Pure and saving removes any stored key and gateway.

Custom OpenAI-compatible endpoint

On the Enterprise plan, select the Custom (OpenAI-compatible) provider to run agents on models Polylane doesn't offer, such as a self-hosted or private deployment. Configure:

• Base URL: the root of your OpenAI-compatible inference API, for example https://your-endpoint.example.com/v1.
• Models URL: optional. A separate endpoint that lists available models. Leave blank to use {base URL}/models.
• Auth header name: the header your endpoint authenticates with. Defaults to Authorization. Use x-api-key or similar if your endpoint requires it.
• Auth header value: sent verbatim as the value of that header, so include any prefix your endpoint expects, for example Bearer sk-.... Encrypted at rest.

Usage on your endpoint is billed by your provider and is not metered against your Polylane plan.

Model tiers

Polylane routes each internal task to a tier. Click Load models to pull the model list from your endpoint, then assign a model per tier:

You must assign at least one of Premium, Standard, Cheap, or Code to save. Unset tiers fall back: Premium, Standard, and Cheap fall back to each other, and Code falls back to them, so setting only Standard runs all agent work on one model. No tier falls back to Code, so don't set Code alone. Embeddings falls back to Polylane's managed embeddings when unset. Reranking is skipped when no reranker is set.

Finish with Test connection. It runs a test generation against your cheapest configured tier.