Integrations

AI Runtime connects to the model providers and MCP servers you protect, to your identity provider, and to the tools your security team already uses. This page lists each integration and what it needs from the network when the gateway is self-hosted.

At a Glance

Model Providers

In gateway mode, the gateway forwards to the providers you add as integrations: OpenAI, Anthropic, Google Gemini, Google Vertex AI, Azure OpenAI, Amazon Bedrock, Mistral AI, Groq, xAI, DeepSeek, Cohere, Together AI, OpenRouter, Perplexity, Fireworks AI, Cerebras, NVIDIA NIM, and a Custom Backend for your own OpenAI-compatible service or Ollama host.

On a self-hosted gateway:

• Allow outbound HTTPS from the gateway to the providers you actually use, and nothing else. In a fully in-app deployment the gateway needs no route to providers at all.
• Provider keys are stored on the integration, encrypted at rest in your own PostgreSQL, and never returned by the API.
• A Custom Backend inside your network is reached over your internal network.

Managed Agents

The gateway also fronts the providers' managed agent platforms: Amazon Bedrock AgentCore, Amazon Bedrock Agent, Azure AI Foundry Agent and Vertex AI Agent Engine. They connect the same way as model providers. See Integrations.

MCP Servers

The gateway sits between your agents and the MCP servers they use, and speaks Streamable HTTP to both. Servers that only offer stdio or the legacy HTTP+SSE transport cannot be connected. See Connect an MCP Server.

On a dedicated gateway, private MCP servers are reached through the private connection to your network: ByteHide resolves their hostnames to their addresses inside the tunnel and declares those addresses on the gateway.

Use Approved MCP servers in Settings to restrict connections further.

Upstream Authentication

Servers that use OAuth sign-in send your browser back to the gateway's /mcp-oauth/callback. Make that path reachable from the browsers of the people who authorize servers. See MCP Authentication.

Identity Providers

For MCP Routes

People calling MCP routes can authenticate with your own identity provider, such as Okta, Microsoft Entra ID or Auth0, instead of a pasted token. You configure it in Settings with its issuer URL. See MCP Authentication.

On a self-hosted gateway:

• The gateway validates tokens against your provider's published keys, so it needs outbound access to your provider's issuer.
• The issuer and its keys URL must resolve to public addresses. Declaring private MCP networks does not extend to identity providers.
• On-behalf-of exchanges each caller's token at your provider's token endpoint on every call.
• Sign in through the gateway registers the gateway's own callbacks as redirect URIs at your provider, so those callbacks must be reachable from your users' browsers.

For the Console

Your team signs in to the console through ByteHide, with TOTP multi-factor authentication, or with Microsoft Entra ID single sign-on over OpenID Connect, set up in your organization's security settings. The console then calls your gateway's control API with the member's session, and your gateway verifies it with the ByteHide platform.

Alert Channels and SIEM

Alerts and Playbooks deliver through the ByteHide platform:

• Notification channels: email and webhooks out of the box, and Slack, Microsoft Teams, Telegram, Discord, WhatsApp and more once connected in your organization's Integrations.
• SIEM destinations: Splunk, Microsoft Sentinel, Elastic, IBM QRadar, Google SecOps (Chronicle), Datadog, Sumo Logic, Wazuh, Grafana Loki, New Relic, Amazon Security Lake, OpenTelemetry, Exabeam, LogRhythm, Logz.io, Cribl Stream and Syslog CEF, configured in your organization's Integrations. Use the Forward to SIEM playbook step to send matching incidences.

SIEM events carry the verdict, detections, identifiers and request context. They never contain prompt or response text, on any destination.

On a self-hosted gateway, alerts are evaluated by your gateway and delivered by the ByteHide platform, so the gateway needs its outbound path to ByteHide. If you close it, alerts are still evaluated and recorded in your database, but nothing is delivered.

Tracing

The gateway exports traces over OTLP and propagates the W3C traceparent header, so its spans join your existing APM traces. Configure it with the standard OpenTelemetry environment variables; without an endpoint, nothing is exported. Each response also carries an x-bytehide-request-id header to match a client-side error to its incidence.

SDKs

On a self-hosted gateway, your applications call your gateway's URL instead of the SaaS one, and in-app applications must reach its verdict endpoints. See Requirements.

Next Steps