CHAPTER I·THE GATE

Every AI request.
One gate.

Gatewyse routes your traffic across every major AI provider with intelligent free-tier chaining, cryptographic audit trails, and a lifetime license you actually own.

See how routing worksRead the architecture

Own it for $2,499 →
EVERY MAJOR AI PROVIDER·One gate
OpenAIAnthropicGoogle GeminiAzure OpenAIxAIMistralCohereDeepSeekAWS BedrockVertex AIGroqTogether AIFireworksPerplexityCerebrasSambaNovaReplicateHuggingFaceAI21DeepInfraLambda LabsCloudflare Workers AIOpenRouterNovitaSiliconFlowNebiusBasetenHyperbolicNVIDIA NIMMoonshotZhipuDashScope (Qwen)MiniMaxOllamavLLMLM StudioLocalAIllama.cppAssemblyAIElevenLabsWhisperComfyUIStability AIVoyageDeepgramCartesiaJinaOpenAIAnthropicGoogle GeminiAzure OpenAIxAIMistralCohereDeepSeekAWS BedrockVertex AIGroqTogether AIFireworksPerplexityCerebrasSambaNovaReplicateHuggingFaceAI21DeepInfraLambda LabsCloudflare Workers AIOpenRouterNovitaSiliconFlowNebiusBasetenHyperbolicNVIDIA NIMMoonshotZhipuDashScope (Qwen)MiniMaxOllamavLLMLM StudioLocalAIllama.cppAssemblyAIElevenLabsWhisperComfyUIStability AIVoyageDeepgramCartesiaJina

CHAPTER II·THE THREE PILLARS

What the gate stands on.

PILLAR I  ·  OWNERSHIP

Own it.

A lifetime license is $2,499. Host it on your cluster, your laptop, or an air-gapped VPC. Bring your own keys. Keep your own audit logs. We never sit between you and your data.

PILLAR II  ·  JUDGMENT

Route it.

Eleven routing strategies, including free-tier chaining that exhausts every provider's free quota before touching a paid key, least-cost, least-latency, failover, weighted, ML-driven task-optimized routing, and a review panel that asks several models at once and synthesises their answers.

PILLAR III  ·  VIGILANCE

Audit it.

SHA-256 hash-chained immutable logs you can verify yourself. Argon2id for passwords. AES-256-GCM with scope-bound AAD on every stored credential. Fourteen guard types. SOC 2-ready by construction.

CHAPTER III·THE JUDGMENT

Eleven strategies. One gate.

Ten of them order the providers into a chain and walk it: Gatewyse checks health, cost, latency, and free-tier balance, and when a provider's quota exhausts or its circuit opens, the gate passes judgment to the next candidate without dropping a packet. The eleventh, review panel, refuses to pick — it puts the question to every model at the top priority level at once and has a synthesizer write the answer they agree on.

CHAPTER IV·THE MODALITIES

Not just chat. Everything.

Thirteen modalities and surfaces. One API key. Route by task, cost, latency, or free-tier balance — the gate decides.

POST /v1/chat/completions

OpenAI-compatible chat with streaming, tool use, and vision.

Providers

  • OpenAI
  • Anthropic
  • Gemini
  • Groq
  • Mistral
  • Together
  • …
curl https://api.gatewyse.com/v1/chat/completions \
  -H "Authorization: Bearer $GATEWYSE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [{"role":"user","content":"Explain routing strategies in one line"}],
    "stream": true
  }'

CHAPTER V·THE VIGILANCE

Security as receipts, not adjectives.

Six laws of the gate. Each one verifiable in code, hardened against the threats that actually matter for an AI gateway: credential exfiltration, audit tampering, provider compromise, and cascading failure.

  1. Hash-chained audit

    Every log entry is sha256(entry ‖ previousHash). Verification recomputes each entry's hash from the record it covers as well as checking the links, so editing a row in place is caught, not just deleting or reordering one — and you run the pass yourself from the admin console. Append-only. Sequence numbers prevent reordering. TTL indexes enforce 365-day retention by default.

  2. Argon2id

    The only password hash we accept. No bcrypt. No SHA-family. Memory-hard by design, resistant to GPU and ASIC attacks. Cost parameters per deployment; defaults exceed OWASP 2024 recommendations.

  3. AES-256-GCM with scope-bound AAD

    Provider API keys are encrypted at rest. An override's scope — the organization or department that owns it — is bound in as additional authenticated data, so a ciphertext lifted from one department's credentials will not decrypt under another's. Ciphertext is version-prefixed for rotation.

  4. Prompt & response guard

    Fourteen guard types run on every request and response: regex PII detectors (SSN, credit card, email, phone, IP, API key), jailbreak and prompt-injection detection, secrets detection, topic and keyword filters, token, rate and cost limits. Toxicity and PII additionally run a local ONNX model — in-process, never an outbound call to a moderation API. Actions: block, redact, mask, or hash.

  5. OIDC + SAML single sign-on

    Okta, Microsoft Entra, Google, and any OIDC/SAML provider. Each organization points at its own identity provider and claims its own domains, so two organizations in one deployment never share a session or a directory. A deployment-wide config is the fallback for the ones that do not.

  6. Provider circuit breaker

    Each provider runs behind a per-instance state machine. After 5 consecutive failures, the circuit opens; routing skips it for 60 seconds; a half-open probe then closes it on success or re-opens on failure. No cascading failures under provider outages.

CHAPTER VI·THE CASCADE

Three business units. One deployment.

Providers and models cascade from the deployment catalogue down to an organization, and from an organization down to a department, by sparse patch. Each level changes only the fields it names and inherits everything else. Nobody clones a configuration to change one line of it.

  1. BASE

    The deployment catalogue

    Every provider the deployment knows how to reach, and every model on it, live in one catalogue. Mutating it requires deployment scope, so a catalogue entry is not something a business unit can quietly rewrite for everyone — and no organization can add a provider the deployment does not have.

  2. ORGANIZATION

    Its own keys, its own directory

    An organization patches a catalogue entry rather than copying it: its own credential, base URL, rate limits, priority, and which models it withdraws or takes up. It points at its own identity provider and claims its own email domains. Reads and writes are bounded by scope, and a refusal is written to the audit trail.

  3. DEPARTMENT

    The narrowest thing that can hold a key

    A department patches its organization's resolved view the same way. Absent is not false — a field you never set keeps tracking the level above, so rotating a key or repricing a model at the top reaches everyone who did not override it, and the ones who did keep what they chose.

And the credentials do not travel with them. The scope that owns an override is bound into its ciphertext as additional authenticated data, so a key one department set will not decrypt in another's context.

CHAPTER VIII·THE SOVEREIGNTY

Buy it once. Or pay it off.

Most platforms rent their gate to you forever. Gatewyse gives you two honest options — own it outright, or amortize the license through a managed deployment and keep it when you're done.

STARTER  ·  PERPETUAL

$2,499once

  • Every future update, forever
  • Self-host anywhere — cluster, laptop, isolated VPC
  • Bring your own provider keys
  • Every provider, every routing strategy, retrieval included
  • Unlimited organizations, departments, users and requests
  • First year of support included
Buy a licence

MANAGED  ·  DEDICATED

$99/mo

Single-node k3s deployment we run for you. Infrastructure and license amortized into one monthly bill.

See managed pricing

CHAPTER IX·THE QUESTIONS

What you're about to ask.

Gatewyse tracks daily free quotas in Redis, per provider and per model. The free-tier-first strategy dispatches to providers with remaining free balance in priority order. When all free quotas exhaust (or circuit-open), it falls back to paid providers automatically. Zero manual fallback logic.
Feature parity. Self-hosted runs wherever you want (Docker, k8s, k3s, bare metal). Managed is our k3s deployment for teams that prefer not to operate it. Switching is a config change plus a data migration script we provide.
Not yet audited. The architecture is SOC 2-ready by construction: immutable audit log with hash chaining, TTL-based retention, Argon2id password hashing, AES-256-GCM at rest with scope-bound AAD, rate-limiter fail-closed behavior, full RBAC. We work with customers on the audit during onboarding.
Yes, and each organization or department can override the deployment's key with its own. Every credential is encrypted with AES-256-GCM, and an override binds the scope that owns it in as additional authenticated data — so a ciphertext lifted from one department's credentials will not decrypt under another's. Keys never leave your deployment on self-host.
Audit logs default to 365-day retention via MongoDB TTL indexes, set per deployment with AUDIT_LOG_RETENTION_DAYS (0 keeps entries indefinitely if you archive them elsewhere). Request/response bodies are not logged by default; you enable that per organization or department if you need it for compliance.
Each provider runs behind a circuit breaker (CLOSED → OPEN after 5 consecutive failures → HALF-OPEN after 60s). Routing skips OPEN providers automatically and probes them in HALF-OPEN. No request lands on a degraded provider.
The licence is bound to the single legal entity that purchased it, not to any individual. A change of ownership or control that keeps that entity intact does not affect it. Spinoffs, subsidiaries operating as separate entities, and restructures that create a new entity need their own agreement — talk to us. Full terms ship with the licence.
Yes. After 36 months of Dedicated, the lifetime license transfers automatically. You can stay on managed, or take the license and self-host while continuing to use infrastructure support at reduced cost.

·EPILOGUE·

You've walked the gate.

Time to stand it up.

Get Gatewyse