Skip to main content
The Barndoor LLM Gateway can sit in front of your Claude Code sessions. Requests run on your own Claude subscription through the gateway, and when a seat hits its rate limit the gateway fails the request over to a fallback model you configure. Developers stay unblocked, and you keep the subscription as the primary, low-cost path while paying for overflow only when you actually spill.

​
How it works

1

Request enters the gateway

Claude Code talks to the gateway using the Anthropic Messages API. Your Claude subscription token is forwarded through to Anthropic (OAuth passthrough), so the request is served on your subscription.
2

Primary target is tried first

The gateway resolves the model to a route. The highest-priority target is your Claude subscription.
3

Failover on a cap

If Anthropic returns a rate-limit error (HTTP 429), the gateway reroutes the same request to the next target in the route, such as an open model, with no action from the developer. The response comes back on the fallback and the session continues.

​
Prerequisites

You need a Barndoor account with admin privileges, the LLM Gateway enabled for your organization, and an active Claude subscription for the developer using it.

​
1. Configure the providers

In the Barndoor web UI, go to Admin > LLM Management > Providers.
  1. Add an Anthropic provider and choose the Claude OAuth auth method. This is the passthrough provider: it stores no API key and forwards the caller’s subscription token.
  2. Add a fallback provider for the model you want to overflow to (for example an open model on a cloud you already use). Give it whatever credentials that provider needs.

​
2. Create a route with failover

Go to Model Routes and create a route under a single model name (the alias Claude Code will target). Add two targets, in priority order:
PriorityTargetPurpose
1 (highest)Anthropic (Claude OAuth)Primary, served on your subscription
2Fallback providerOverflow when the subscription is rate-limited
A model route with two priority-ordered targets, a primary and a fallback
Targets are tried top to bottom. Set the primary’s retry-on-429 count if you want the gateway to briefly retry the subscription before spilling; leave it at 0 to fail over immediately.

​
3. Create a gateway API key

Go to API Keys and create a key. This authenticates Claude Code to the gateway; it is separate from your Claude subscription token, which the gateway forwards to Anthropic.

​
4. Point Claude Code at the gateway

Stay logged into Claude Code with your Claude subscription so it sends its OAuth token, then set these environment variables and launch claude: 
# Do NOT set ANTHROPIC_API_KEY — that switches Claude Code to API-key mode
unset ANTHROPIC_API_KEY

export ANTHROPIC_BASE_URL="https://<your-gateway-host>/api/llm-gateway"
export ANTHROPIC_CUSTOM_HEADERS="x-api-key: <your-gateway-api-key>"

# Map Claude Code's model slots to your route's alias
export ANTHROPIC_DEFAULT_OPUS_MODEL="<your-route-alias>"
export ANTHROPIC_DEFAULT_SONNET_MODEL="<your-route-alias>"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="<your-route-alias>"
The split is the key detail: the x-api-key header authenticates you to the gateway, while Claude Code’s Authorization: Bearer subscription token is what the gateway forwards to Anthropic.

​
5. Verify

Run a normal Claude Code request. It should complete on your subscription. To see the failover, drive the seat until Anthropic rate-limits it: the gateway logs the 429 and reroutes to the fallback, and the request still returns a completion.
Subscription terms. Provider subscriptions are intended for individual, interactive use. Route each developer’s own seat and keep usage interactive. Pooling seats behind a shared gateway, or driving them as shared automated capacity, may violate the provider’s terms of service. Review your model provider’s terms before enabling this pattern, especially at scale.