Swap one URL. Govern every AI request.
Point your SDK's base URL at AegisPlane and keep building. Every request then flows through 9 inline stages (auth, policy, threat protection, PII redaction, routing, cost control, inference, rehydration, audit) before the response reaches your users.
Watch a request flow through, end to end
This is the real pipeline. Click any step to zoom in and see what it does.
Your AI App
API call →
LLM · Model Execution
OpenAI
Anthropic
Ollama
Groq
Gemini
Mistral← secured response
↑ click any gateway block to see what happens inside
Explore each stage
Open any stage to see exactly what it checks, why it matters, and what it returns.
Only authorized requests get through
Identity and access control happens first, before any other processing. Invalid requests are rejected instantly.
- Supports API key, JWT, OAuth, and mTLS out of the box
- Tenant and permission resolution in a single step
- Unauthorized traffic is rejected before touching any resource
- Every access attempt is logged and auditable
Governance enforced before the model sees anything
Eight frameworks are checked in real time on every request. On a violation the gateway blocks or warns (you set the action per policy) before inference starts.
- EU AI Act, NIST, GDPR, HIPAA, ISO 42001/27001, SOC 2 and internal policy, evaluated in parallel
- Dynamic risk classification adapts to each request context
- Violations are flagged in real time: block or warn per your policy
- Every check produces traceable, exportable evidence
Stop hostile inputs and abuse patterns
Every prompt is scanned for injections, jailbreaks, and data exfiltration attempts. Threats are blocked in milliseconds.
- Detects prompt injections, jailbreaks, role-play bypasses, and exfiltration
- Regex, ML-based detection, and configurable rulepacks per tenant
- Block instantly or log as warning. Your choice per threat category
- Each threat event is recorded with full context for tuning
Sensitive data never enters the model
PII is automatically detected and masked before the LLM processes the request. User data stays private.
- Detects names, emails, SSNs, IBANs, credit cards, and medical IDs
- Fully configurable per tenant, customer, and use case
- Redacted before any provider, rehydrated on response: request-scoped, never stored as durable tokens
- Every detection is logged with entity type and risk level
Route to the best provider per request
AegisPlane picks the optimal model based on latency, cost, quality, and health. Fallback is automatic.
- Automatically routes to the fastest, healthiest provider in real time
- Switches to a backup provider the moment one degrades
- Add or remove providers without touching your code
- Every routing decision is logged and explainable
Spend is controlled before it happens
Budgets and rate limits are enforced pre-execution. Requests are rejected if they would exceed limits, not after.
- Set budgets per tenant, provider, or model. As granular as you need
- Requests are blocked before the provider is ever called
- Alerts fire at 80% usage so you act before limits are hit
- Full real-time spend visibility across every team and model
The model only runs on clean, approved requests
LLM inference happens only after all 6 upstream controls pass. No unauthorized, unsafe, or over-budget request ever reaches this step.
- Inference runs only on requests that passed every prior control
- No unauthorized, non-compliant, or over-budget request reaches the model
- Model errors are caught and handled gracefully by the gateway
- Every inference event is recorded for cost tracking and audit
Restore PII safely in the final response
After the model responds, masked data is restored only when it's safe and the response passed all checks.
- Masked values are restored precisely where they were in the original
- Rehydration only runs if the response passed all output checks
- Configurable per tenant, customer, and response type
- Users get a complete, useful response. Privacy intact throughout
Every decision is logged and traceable
From authentication to output, every step is recorded with full context. Ready for auditors and investigators.
- Structured event types across all pipeline stages [PLACEHOLDER: confirm exact count, current copy says 12]
- Filter by tenant, customer, event type, or severity in seconds
- Full request traces correlated end-to-end via OpenTelemetry, exported to your stack
- One-click export for EU AI Act, SOC 2, and ISO 27001 reviews
Point one URL. Keep your code.
Swap your provider's base URL for your AegisPlane endpoint. Your OpenAI, Anthropic, or any OpenAI-compatible SDK keeps working unchanged. No rewrites, no new dependencies, no agent to deploy.
base_url = "https://api.openai.com/v1"base_url = "https://api.aegisplane.com/v1"One endpoint across 11 providers. Works with the OpenAI, Anthropic, and any OpenAI-compatible SDK. Streaming and function-calling pass through unchanged.
Run it wherever you need
Same gateway, same controls. You choose where it lives.
Managed Cloud
We host, scale, and patch it. You get an endpoint and start sending traffic, with nothing to operate.
- Get an endpoint and an API key, start sending traffic
- Automatic updates, scaling, and provider fallback
- Low-latency, redundant by default
Self-hosted
Run the gateway in your own VPC or on-prem. Deploys like infrastructure (Helm/Terraform, config-as-code) so data never leaves your network.
- Your infrastructure, your network, your keys
- mTLS, RBAC, and OTel built in
- SPIFFE/SPIRE mTLS by default
Governance without the latency tax
All nine stages run inline on every request, in parallel where possible, to keep added overhead low. If the gateway degrades, it fails open, so a gateway issue never becomes an app outage.
How the flow works in practice
The questions teams ask before pointing production traffic at the gateway.
It's built for high availability and fails open. A gateway problem must not take your app down. On managed cloud it runs redundant, auto-scaled instances; self-hosted, you run your own replicas. If an upstream provider fails, requests fall back and retry automatically.
The gateway processes each request in-memory to apply policies, then forwards it. PII is detected and redacted before any provider sees it, then rehydrated on the response, request-scoped, not stored as durable tokens. You control what's logged: full traceable audit evidence, or metadata only. Self-hosted, data never leaves your network.
Yes. Streamed token responses pass through the same governance stages without breaking the stream, so latency-sensitive and chat experiences keep working as-is.
No. You point your provider's base URL at your AegisPlane router and keep using the OpenAI, Anthropic, or any OpenAI-compatible SDK exactly as before. No new dependencies, no rewrites.
Low, measured overhead. All nine stages run inline and in parallel where possible, so governance doesn't become a latency tax on your users, and the gateway fails open if it ever degrades. [PLACEHOLDER: publish a measured p50/p95 figure here once benchmarked.]
Yes. Policies, guardrails, PII rulepacks, routing, and budgets are all configurable per tenant, per customer, and per risk level: block, warn, or allow as your context requires.
Ready to get started
Deploy it, swap one URL, keep building.
Every stage runs inline and automatically. No code changes, no SDK swap, no agent to deploy, and you can self-host it in your own VPC.