LLMGuard
Every AI. One secure gateway.
Centralize, secure and monitor every AI model your company uses.
One API for every major provider
AI adoption outpaced governance.
Every team picked its own AI tools. IT inherited the consequences.
Shadow AI & key sprawl
Dozens of provider keys in scripts, CI and laptops. No one knows which application uses what.
Sensitive data leaving
Customer emails, IBANs and credentials pasted into prompts and sent to external providers.
Unbounded spend
Fragmented invoices, no budgets, no quotas — the first signal is the end-of-month bill.
One API and one control plane for every AI provider.
Stop sensitive data before it leaves your organization.
Monitor usage, enforce quotas and control AI spending.
Swap provider keys for one LLMGuard API key.
Point any OpenAI-compatible SDK at the gateway. Requests are routed to OpenAI, Anthropic, Gemini, Mistral or DeepSeek with your own keys, encrypted at rest.
Bring your own keys
Provider credentials are encrypted with AES-256-GCM and never leave the server. You keep your provider contracts and pricing.
Provider abstraction
One OpenAI-compatible endpoint, streaming included. Anthropic and Gemini are translated transparently.
Teams & applications
Attribute every request to a team, an application and a key — the foundation of chargeback and governance.
Sensitive data never reaches the model.
Every prompt is inspected before it leaves your network. Block, mask or log — per organization, team or application.
Summarize the complaint from claire.martin@client.fr about invoice CUS-48213377. Card on file: 4970 1012 3456 7890.
Summarize the complaint from [REDACTED_EMAIL] about invoice …
Built-in detectors
Emails, phones, IPs, cards (Luhn), IBAN (mod-97), French NIR, API keys, JWTs, AWS keys, private keys, secrets.
Mask, don't just block
Replace values with [REDACTED_EMAIL] and keep the request useful.
Custom rules
Your customer numbers, project codenames — with ReDoS-safe regex validation.
Zero-value logs
Detections record what and where — never the sensitive value itself.
Know who consumes what — and cap it.
Real-time visibility by provider, model, team, application and key, with quotas enforced on every request.
Monitoring
Requests, tokens, latency, error rates and estimated cost — with a request explorer that stores metadata only.
Budgets
Monthly budgets per team or application, forecasts, and alerts at 80% and 100%.
Quotas
Requests/minute, tokens/day, cost/month… block or alert, enforced atomically across gateway instances.
Two lines of change. Zero lock-in.
The gateway speaks the OpenAI Chat Completions protocol, including streaming. Errors are normalized; every response carries a request ID.
from openai import OpenAI
client = OpenAI(
base_url="https://gateway.llmguard.io/v1", # ← LLMGuard
api_key="llmg_live_…", # ← one LLMGuard key
)
client.chat.completions.create(
model="anthropic/claude-sonnet-5-5", # or openai/…, google/…, mistral/…
messages=[{"role": "user", "content": "Draft the release notes"}],
stream=True,
)Built like infrastructure, not like a chatbot.
Simple pricing, announced soon.
Bring your own provider keys — we never mark up AI usage. Early access is free while we finalize plans.
- Up to 3 providers
- DLP baseline
- 30 days of history
- Unlimited providers & teams
- Custom DLP rules & policies
- Budgets, quotas, webhooks
- SSO / SAML & SCIM (roadmap)
- Self-hosted gateway
- Data residency
Questions IT teams ask first.
Do you store our prompts?
No. By default only metadata is stored (who, which model, tokens, cost, DLP outcome). Content logging is an explicit, encrypted, audited opt-in.
Do we have to change our code?
Only the base URL and the API key of your OpenAI-compatible SDK. Models are addressed as provider/model.
Who pays the AI providers?
You do, directly, with your own keys (BYOK). Costs in the console are estimates from a central price catalog.
What happens when a quota is reached?
Depending on the policy, requests are rejected with HTTP 429 QUOTA_EXCEEDED, or allowed while an alert is raised.
How fast is the gateway?
Authentication, DLP, policies and quotas run on in-memory snapshots plus one or two Redis round-trips — typically single-digit milliseconds.
Can we self-host it?
The architecture separates the control plane from the gateway so it can run in your own infrastructure. Self-hosting is on the enterprise roadmap.
See every AI request in 5 minutes.
Connect a provider, create a key, send a request — it shows up instantly.
Start for free