Providers¶
The Providers page (admin console) is where you connect the LLM model providers that power every chat, agent run and background task. Platform providers are the only model source for users — there is no user-facing BYOK, so a model is only selectable once an admin has configured its provider and enabled the model here.
The core flow is simple: set a key per provider — its available models are then fetched from the provider API. Toggle which models users may pick.
Key storage must be enabled
Providers can only be configured when a master key is present. Without one,
Add provider is disabled and a warning banner is shown — keys cannot be
stored.
Adding a provider¶
Click Add provider and pick one from the menu. The picker lists every
provider from the static catalog that isn't configured yet, plus a
Custom provider (self-hosted) option for a DB-defined endpoint.
The add dialog asks for:
| Field | Notes |
|---|---|
Display name |
Only for providers that require a base URL. |
Base URL |
For OpenAI-compatible / custom providers, e.g. https://…/v1. |
API key |
At least 8 characters. Stored encrypted; only the last 4 digits are shown afterwards. |
Save & fetch models stores the key and immediately queries the provider API
for its model list. A toast reports how many models were found (or warns if none
were — check the key/base URL).
Custom & keyless providers¶
The Custom provider dialog defines a DB-driven provider without a code edit:
a Provider ID (lowercase letters, digits, -, _), a Display name, a
Base URL, an optional Default max tokens, and a Keyless toggle for local
servers that need no key (e.g. a self-hosted Whisper/TTS endpoint). Keyless
providers are reached directly at their base URL and are usable immediately.
Managing a configured provider¶
Each configured provider expands to show its key, governance settings and model
list. The header summarizes how many models are active (e.g. 4/12 models
active).
Replace key— rotate the API key (also re-fetches models).Refresh models— re-query the provider's model list.Remove— delete the provider, its key and its model selection.
Governance¶
Two provider-level governance controls apply to all the provider's models:
| Field | Effect |
|---|---|
Trust tier |
Data-governance clearance — internal (on-prem) reaches the most sensitive data, regulated is for compliant external providers, unregulated is any external provider. This is the fail-closed data-classification gate, not a routing hint. |
Routing tags |
Category tags (frontier, coding…) that drive Auto model selection. Not governance. |
Choosing which models users may pick¶
Below governance, each discovered model has a toggle. Only enabled models are selectable by users. When a model is enabled, three per-model settings appear:
| Control | Effect |
|---|---|
| Model tags | Per-model routing tags (frontier, coding…) layered on top of the provider tags, used by Auto selection. |
Quality (0–20) |
An admin within-tier ranking nudge for Auto — breaks a capability tie toward this model (e.g. a direct provider over a reselling proxy). It never crosses a trust tier or overrides the capability score. |
| Pricing (tag icon) | Opens the per-model pricing override dialog. The icon is highlighted when an override is set, grey when the genai-prices default applies. |
Model pricing¶
Pricing overrides are edited inline from each model's tag icon (the standalone pricing page is gone). The dialog sets prices per 1,000,000 tokens in the chosen currency:
| Field | Notes |
|---|---|
Input / Output |
Prompt and completion price per Mtok. |
Cache hit / Cache write |
Prompt-cache read/write price per Mtok. |
Currency |
Defaults to USD. |
Remove override reverts the model to the catalog price.
Overrides vs. the catalog
These prices override the built-in catalog (genai-prices). Anything not set
here is priced via genai-prices automatically. An override with an empty
provider applies model-wide; one pinned to a provider takes precedence over a
provider-less row for the same model.
The page also hosts related admin settings that depend on the same provider list — the safety guard model, the memory curator model, voice (STT/TTS) models, org governance minimum trust tiers, and an append-only governance audit log of provider/model tag changes.