Model availability

Why a model is missing from the picker, or why requests to it are refused. Model access is decided by the organization's allowlist, per-device overrides, which providers are configured, and the plan.

The model is not on the organization's allowlist

Admins choose which gateway models the organization exposes under Manage → Settings → Available models (shown when "Use Harriet as LLM proxy" is on). A model unchecked there does not appear in anyone's picker. See Available models.

One wrinkle by plan: on the free tier, "All models exposed for your configured LLM keys are available on this plan" — a stored allowlist is ignored, and the console notes that "Choosing which LLM models are available requires Endpoint AI Team or higher." So on Free, a missing model is never an allowlist problem; look at providers and plan gating below instead.

A device-level override is in effect

Each device can carry its own model list. On the device detail page, the Available models section shows either "Using account default." or "Custom model list for this device." with an Inherit from account link. If one machine is missing models the rest of the organization sees, check for a custom list here and either add the model back or select Inherit from account.

The provider is not configured

A model only appears when Harriet can actually reach its provider. That requires, per provider (OpenAI, Anthropic, Google), one of:

• An organization BYOK key stored in Provisioner settings, or
• A managed credits route, where Harriet supplies the upstream key and bills prepaid credits.

If neither exists for a provider, its models are absent from the picker, and a request sent anyway is refused with 401 and a message saying no key is available for that provider. Managed calls additionally require a positive credit balance; when credits run out the proxy returns 402 with insufficient_credits — see Error codes.

Anthropic (EU) models need the right plan

The Anthropic (EU) group (Opus 4.7 EU, Haiku 4.5 EU) is gated by plan: "Anthropic (EU) models require Endpoint AI Business or higher." These models are Harriet-managed resale only — there is no BYOK for them, so they are billed through managed credits, and traffic stays on EU infrastructure. Details in EU data residency and Plans.

Editing which models are available (organization or per-device) also requires Endpoint AI Team or higher.

The device is refused: device_llm_access_denied

When a desktop app calls Harriet's LLM gateway from a device that is not allowed to use it — typically a device configured only for Claude Desktop with Claude account, where people use their own Claude accounts and Harriet supplies no models — the proxy answers 402 with error code device_llm_access_denied:

This device is not allowed to use Harriet LLM access.
Update the device to "Claude Desktop (no Claude account)" in the Harriet console.

This usually means the desktop app is still pointed at the gateway from an earlier configuration. Fix it in the console: open the device detail page and, under Allowed desktop providers, enable the surface the person should use — Claude Desktop (no Claude account) for gateway-supplied models, or Harriet Desktop. Only providers enabled in account defaults can be turned on per device. How the gateway routes requests is covered in LLM proxy.