Frequently asked questions

What is Ramp Router's data retention policy?

Router records model inputs, outputs, and tool calls for debugging, semantic analysis, and product improvement. Recorded content is retained for one year by default, and Ramp removes detected personally identifiable information before using that content to improve the product.

You can opt out of content recording at any time in Settings. The preference applies account-wide to your self-service Router API keys and propagates asynchronously, so requests may briefly use the previous setting. Opting out stops future content recording after that propagation window; it does not delete existing archives, and Router may still retain operational metadata such as model, provider, usage, cost, latency, and request status.

Please see the Ramp Router Privacy Policy for information on how Ramp manages personal information.

Where are the models hosted?

All models run on U.S.-based infrastructure. Ramp does not host the model weights. Models run in the infrastructure of their underlying provider, such as OpenAI, Anthropic, Google Vertex AI, Fireworks, or xAI. Router is the gateway between your application and those providers: it selects a model, sends the request to that provider, and returns the response in a consistent API format.

The processing region and data-residency options depend on the selected provider and model. You can see which provider handled a request in Logs.

What API format does Router use?

Ramp Router supports two API surfaces:

• POST /v1/responses - OpenAI Responses API format
• POST /v1/messages - Anthropic Messages API format

The OpenAI chat completions endpoint (POST /v1/chat/completions) is not supported. If your client uses the chat completions format, it will need to be updated to use the Responses API before connecting to Router.

Both surfaces route to the same underlying providers. Responses requests always return an OpenAI Response object; Messages requests return the Anthropic Messages shape.

How does Switchyard routing work?

Switchyard is an opt-in routing strategy for coding-agent traffic (Claude Code, Codex, and similar harnesses). For eligible requests, it extracts signals from the conversation - error severity in recent tool outputs, tool-call categories, test results, and turn depth - and uses those to decide whether to route to a capable frontier model or a cost-efficient model. Simpler tasks route down; complex or error-heavy turns route up.

Requests pass through to the requested model unchanged when Switchyard is not enabled for your account, the request includes image input, the requested model is already in the cost-efficient tier, or an eligible cost-efficient model is unavailable.

You can see which model handled each request in Logs. Routing strategy settings are configurable in your dashboard.

Does Router support bring-your-own API keys (BYOK)?

Yes, for OpenAI, Fireworks, and xAI. You can add your own provider credentials in Settings, scoped per provider. When a request is served using your key, the provider bills you directly and Ramp Router does not charge you for that usage.

If a request fails on your key due to an auth error, rate limit, or provider error, Router can optionally fall back to its shared key and bill normally. You can disable this fallback in Settings if you want strict BYOK-only behavior.

BYOK is not currently available for Anthropic or Google Vertex AI models; those use Router's shared credentials.

Why was my API key deactivated?

Why a key gets deactivated. Requests are screened against usage policy categories. A request that scores above the threshold in a category deactivates the key that sent it.

Why several strikes can land at once. Screening runs asynchronously, and coding agents retry automatically. A single prompt that is retried several times can register as several separate violations before you see the first one.

What a ban means. Repeated violations lock the account, and new keys can't be created until an administrator lifts it.

Getting unlocked. Contact router@ramp.com. Bans can be lifted by an administrator, and both the ban and the lift are recorded in the audit log.

If you think it was wrong. Say so when you contact us. Screening scores are recorded per request, so a decision can be reviewed rather than argued about.