Public betaOpen source. Built in the open.Explore the beta
Operate the control layer

Understand what you run.

Start with security and deployment, connect providers through the gateway, then add policy-aware tool and agent traffic.

Updated · Public beta

Start with one governed request

The quickstart creates an auth-enabled local gateway and an OpenAI-compatible client. Until the release is published, it uses a pinned source build rather than a nonexistent download.

Configuration & hot reload

Use GatewayConfig with gateway.auth, gateway.cache_scope, and gateway.log_payloads. File-driven routes and supported policy settings hot-reload; process and storage settings can require a restart.

Read the canonical guide

Authentication & gateway keys

Enable OMNISWITCH_AUTH=true. Use OMNISWITCH_BOOTSTRAP_API_KEY for first-run access, then create a personal gateway API key. Roles and workspace scope govern access. OIDC workload identity is separate from dashboard SSO.

Read the canonical guide

Routing & providers

Configure the providers[] list and the routes map of named routes. Native adapters and custom OpenAI-compatible endpoints share fallback, retry, timeout, weighted routing, and CEL condition controls.

Read the canonical guide

Local guardrails

Configure guardrails.actions, guardrails.rules[], and opt-in webhooks. Deny and redact are enforcement actions; checks are not a guarantee against every attack. Output stream buffering can delay delivery.

Read the canonical guide

Budgets & limits

Gateway API keys have budgets; rate_limit.requests and rate_limit.window control request windows. Optional rate_limit.redis_url shares windows across replicas. Token/USD policies and per-key TPM/USD-day limits are part of the unreleased v0.2.0 work.

Read the canonical guide

Provider credentials

Provider credentials are upstream secrets, not gateway API keys. Reference secrets through providers[].api_key_env. Never embed a production credential in checked-in YAML, examples, traces, or screenshots.

Read the canonical guide

MCP tool governance

Beta

MCP is Beta. Configure mcp.enabled and mcp.targets[] for HTTP/SSE or persistent stdio targets. Federation and policy-gated calls are available. Per-tool discovery filtering and configurable RFC 8693 token exchange are prepared for v0.2.0; OAuth protected-resource discovery is deferred.

Read the canonical implementation

Governed A2A model access

Beta

SendMessage, SendMessageStream (SSE task events), GetTask, best-effort CancelTask; tasks in memory; no push notifications; no downstream agent proxy. This interface uses the governed model pipeline; it is not a durable orchestration service.

Read the canonical implementation

Observability

Inspect request metadata and guardrail events in the gateway. Configure observability.prometheus_enabled and OTLP settings for your own monitoring stack. Payload logging stays an explicit opt-in. The redesigned operator console is part of the unreleased release work.

Read the canonical guide

Deployment

Start with a binary or container. Use one replica with SQLite. Multi-replica v0.2.0 deployments require Postgres; Redis shares rate-limit windows. Configure authentication, TLS termination, backups, and retention before exposing a gateway.

Read the canonical guide