Guardrails
LLM guardrails for every API key you hand out
Someone on your team will paste a customer's email into a prompt. Put AI guardrails on the key and Hopscotch blocks it before any model sees it.
- Blocked requests cost $0
- Provider list price with no markup
- We never store your prompts or responses
- One OpenAI-compatible API, 500+ models
Production
Questions people ask before they turn this on
What are LLM guardrails?
Rules a request has to pass before it reaches a language model. On Hopscotch you attach them to an API key, and the gateway checks the budget, zero data retention and sensitive data. Fail one and the request is blocked at no charge.
How much do guardrails cost?
Nothing. Guardrails are free. You pay provider list price for requests that go through and nothing for blocked ones. Per-model rates are on the LLM API pricing page.
Do blocked requests cost anything?
No. You pay nothing for a blocked request, whichever check stopped it. The request logs record the reason, so you can fix whatever caused it.
Why put guardrails at the gateway instead of in the prompt?
GenAI guardrails in a system prompt or one app's code leave gaps. A model can ignore an instruction, and an app check misses a teammate's script. LLM gateway guardrails cover every request on the key, from any app, alongside the guardrails for LLM apps you already run.
Can I use these as guardrails for AI agents?
Yes. AI agent guardrails work like any other key: give each agent its own key and guardrail. A looping agent hits its budget and stops, and one that grabs a secret as context gets blocked.
Which checks should I turn on first?
Start from what the key is for. Prototypes and internal tools usually need only a budget. Anything that touches customer data should add the sensitive-data check and zero data retention. You can change a guardrail later without touching code.
Should guardrails for AI in production differ from testing?
Usually, yes. Give production and test traffic separate keys, each with its own guardrail. Test keys can run on a small budget with looser rules, while production keys carry the checks your customers' data needs.
What does a blocked request return?
A 403 error that names the check, for example 403 sensitive data match. Every guardrail block returns 403, so handle it like any other API error.
Does anyone at Hopscotch read my prompts?
No person reads them. Our software checks each prompt in transit, and we don't store prompts or responses.
Is my prompt sent anywhere else to be checked?
No. The sensitive-data check runs inside Hopscotch and adds about 1 ms. Your prompt isn't sent to another model or company to be classified.
What counts as sensitive data?
Email addresses, IP addresses, secrets, API keys and passwords. Passwords have no set shape, so the check catches common formats and can miss unusual ones. We block the request rather than quietly edit your prompt. Remove the value and send it again.
How is a guardrail spend cap different from a per-key spend limit?
A per-key spend limit is a lifetime total, like $500 until the key expires. A guardrail cap works like a rate limit and resets each period, like $50 a day. If both are set, the guardrail takes precedence.
Can a routing profile override a guardrail?
No. A guardrail lives on the key and applies to every request on it, whichever models the routing profile lists.
What does Hopscotch store?
Only metadata: model, provider, token counts, cost, latency and outcome, kept for 90 days by default. We never store prompts or responses, including playground runs. API keys are kept as a one-way hash, so even we can't read them back.
How is this different from OpenRouter guardrails?
OpenRouter covers more checks today. However, Hopscotch covers budgets, zero data retention and sensitive data, with no fee on top of provider list price. The full comparison is on our OpenRouter alternative page.
Put a guardrail on your next key
Create a key, attach a guardrail and keep your endpoint. Blocked requests are free.