Skip to main content
Self-hosted. You run the gateway container inside your own network (VPC, on-prem, or even a laptop). Your applications call your gateway, not Guardway. The dashboard at app.guardway.ai is SaaS — it configures and observes gateways but never proxies inference traffic.
No. The gateway forwards requests directly to the provider you configured (OpenAI, Anthropic, etc.). Prompts and completions stay on the gateway — only aggregate telemetry (token counts, latency, cost) and administrative audit logs flow to the cloud dashboard. See Privacy and Security.
20+ presets out of the box — OpenAI, Anthropic, Google Gemini, Groq, Mistral, DeepSeek, xAI, Perplexity, Together, Fireworks, OpenRouter, Cohere, AWS Bedrock, Azure OpenAI, HuggingFace TGI, NVIDIA NIM, Ollama, LM Studio, vLLM, plus specialty providers (Voyage, AssemblyAI, ElevenLabs, Fal.ai). Anything else that speaks the OpenAI API works via the Custom preset. See Connect a provider.
Today they are the same keys, issued from Configuration → API keys and valid on every gateway in your organization. Per-gateway scoping is on the roadmap. See API keys.
Guardrails run on the gateway before the request reaches the provider (and can also inspect the response). Built-in checks cover PII, hate speech, prompt injection, keyword lists, and IP allow/block-lists. Detection uses small language models (SLMs) — see Guardrails.
Yes. Register MCP servers in Configuration → MCP and choose which servers and tools each API key or application can use. The gateway can combine several MCP servers behind one URL, and no LLM provider is required. Tool calls are logged separately — see MCP and Logs → MCP logs.
Yes. You can register as many gateways as you want — production, staging, per-region, etc. Providers and models are attached per-gateway, so each gateway holds its own credentials and model inventory. Usage dashboards roll up across all of them. See Gateways.
It needs:
  • Outbound HTTPS to api.guardway.ai (control plane).
  • Outbound HTTPS to the LLM providers you use.
  • Inbound access on port 8080 from whoever calls it.
No inbound from the internet is required. See Requirements.
Routing rules support fallback strategies — next-priority, lowest-cost, lowest-latency, fail. You can chain fallbacks so a single provider outage doesn’t break your traffic. See Routing.
Set them at three levels: per API key, per team, or per routing rule. Spend budgets come in three kinds: per API key (a lifetime cap that never resets), per application (daily, monthly, or lifetime, shared by the application’s keys), and per team (calendar month). Requests are blocked when a cap is hit. See Budgets and Alerts for how they compare and which alerts each sends. Separately, a routing rule can switch to a fallback provider when its own budget limit is reached.
The public management API is on the roadmap. The gateway itself already exposes a stable OpenAI-compatible inference API today. See API reference.
Latest Chrome, Firefox, Safari, and Edge. Internet Explorer is not supported. See Limitations.
Bugs: email support@guardway.ai with repro steps. Security: email security@guardway.ai privately — do not open a public issue. See Support.