Guarrix — questions developers ask

Guarrix is a security layer that sits between your application and your AI provider. It checks the messages you send to AI, and the replies that come back, for private data and manipulation attempts — without storing the content itself.

If your app sends real user data to an AI model — names, emails, support tickets, internal documents — you're relying on that model, and everyone who built it, to handle that data responsibly. Guarrix adds a check in front of that, so you don't have to trust the AI provider alone.

No. Requests and responses are processed in memory only and discarded right after. The audit log records metadata — timestamps, token counts, and what type of data was found — never the content itself.

No. We never see your content beyond the moment we check it. Nothing you send through Guarrix is used to train or fine-tune any model, ours or anyone else's.

No. Guarrix sits in front of the provider you already use — OpenAI, Anthropic, Mistral, EUrouter, OpenRouter, Google Gemini, xAI Grok, Hugging Face, or any OpenAI-compatible endpoint. You keep your own API key and your own provider relationship.

Encrypted with AES-256-GCM at rest. The encryption key lives in a separate environment variable that never touches the database. The plaintext key is only held in RAM during a request and is never logged.

Names, email addresses, phone numbers, IBAN and credit card numbers, IP addresses, passport and national ID numbers, dates of birth, and more. You can turn detection types on or off per tenant.

Yes. The guardrail check itself runs in the EU on Hetzner infrastructure — see "Where is my data processed?" below for how this applies when your AI provider is outside the EU. Dot-0 is the data controller under Dutch law. You retain all rights under GDPR Articles 15–22, including the right to access, correct, and delete your data.

It's text hidden inside a normal-looking message that tries to make an AI ignore its own rules — for example, telling it to reveal information it shouldn't, or to act outside the boundaries your application sets. This is often called prompt injection. Guarrix checks for it on every request.

Depending on how you've configured the rule, Guarrix blocks the request, removes the sensitive part, or flags it and lets it through. Either way, your application gets a clear result back — never a silent failure.

Yes, by default. Prompt injection detection and the built-in private-data checks are active on every plan from the start. You can adjust what gets blocked, redacted, or just flagged from your dashboard.

Yes. Every reply from the AI is checked before it reaches your application — for the same private-data types and for content that violates your outbound rules. If something doesn't pass, it's blocked or scrubbed, not forwarded.

The checking itself runs on our own servers in Europe, hosted with Hetzner, with no US parent company involved. What happens after that depends on your AI provider — if you use a US-based model, your message still travels there for the AI reply itself. Guarrix's part of the job stays in the EU either way.

Create a free account, add your LLM API key in the dashboard, point your application at https://api.guarrix.com/v1/chat/completions, and use your Guarrix key instead of your LLM provider key. The free tier includes 500 requests per month.

Guarrix offers five plans: Free (500 requests/month), Starter (€59/month, 5,000 requests), Pro (€199/month, 25,000 requests), Business (€599/month, 100,000 requests), and Enterprise (custom volume with dedicated infrastructure, pricing on request). Overages are charged per request above your plan's included volume.

Yes. The free plan includes 500 requests a month, no time limit, and no credit card required. You can upgrade whenever you're ready, directly from your dashboard.

Yes. You can cancel from your dashboard settings at any time. Your subscription continues until the end of the current billing period. Consumer customers in the EU have a 14-day withdrawal right on new subscriptions under Dutch consumer law.

Requests above your plan's included volume are charged at the overage rate shown on the pricing page. You can set a monthly spend cap in your dashboard to prevent unexpected charges.

The built-in entity types (name, email, phone number, and so on) are a fixed list — you can't add a new one to that list. For anything else, like a password field or an internal project code, use Rules: define a regex pattern or keyword list, give it your own label, and choose whether to block, redact, or allow matches. Rules are configured per tenant from the dashboard.

The check itself typically adds a small amount of latency — usually well under what you'd notice next to the AI model's own response time, which is almost always the slower part. If speed is critical for your use case, test it against your own traffic before going live.

No. The built-in checks work out of the box with sensible defaults. Custom rules are optional — most teams never need to touch them — and everything is managed from a plain dashboard, not a configuration file.

Yes, as long as your model exposes an OpenAI-compatible API endpoint. Point Guarrix at your endpoint URL instead of a cloud provider. Your model endpoint must be reachable from Guarrix's infrastructure over HTTPS.

Business and Enterprise plans include a 99.9% uptime SLA. Free and Starter plans target the same infrastructure but without a contractual SLA. Service status is published at status.guarrix.com.

Email hello@guarrix.com. Business and Enterprise customers have access to a priority support queue with a 4-hour response guarantee during EU business hours.

From your dashboard you can export your audit log as CSV or JSON for any date range up to 90 days. Exported data includes request metadata — timestamps, token counts, PII labels — but never prompt or response content.

Guarrix works with OpenAI, Anthropic, Hugging Face, Mistral, EUrouter, OpenRouter, Google Gemini, and xAI Grok, plus any self-hosted or private model that exposes an OpenAI-compatible endpoint. Chinese model providers (e.g. DeepSeek, Qwen, Zhipu, Moonshot) are not currently supported.

Guarrix doesn't intercept general web traffic — that's a different kind of product (a network firewall or secure web gateway). Routing usage through Guarrix by design, rather than by network trickery, typically means two things on your side: giving employees a sanctioned path (Guarrix's own chat interface) and restricting direct access to the consumer AI sites at your network level (DNS filtering, firewall rules, or device policy on managed laptops). Any restriction on how employees can use their work devices may need consultation with your works council, employee representatives, or legal/HR team under local law — this isn't something we can advise on, so check with them before implementing.