Control who gets access.
Issue API keys with per-key budgets. Scope access with roles and workspace boundaries.
Hashed keys · budgets · rate limits
Self-hosted AI. Security in your hands.
Your AI.
Under your
control.
Secure and route AI traffic through a gateway you run. Keep keys, local guardrails, budgets, and request traces in your own infrastructure.
Apache-2.0 · No cloud control plane
ONE REQUEST. YOUR RULES.Architecture
Your app / OpenAI SDK
Your infrastructure
OmniSwitchSelf-hosted gateway
Authenticate
Guard
Route
Observe
auth: truecache_scope: api_keylog_payloads: false
Cloud providersNative adapters
Local modelsOpenAI-compatible
Own the control layer.
Keep the deployment simple.
One Go binaryBinary, Docker, or Kubernetes
SQLite built inRedis optional for shared rate limits
Apache-2.0Inspect it. Run it. Make it yours.
01 / Secure self-hosted control
Set the boundaries.
See what crosses them.
Policy, visibility, and ownership where your AI traffic meets your infrastructure.
Check what passes through.
Run local PII, injection, and secret checks. Configure deny and redact actions.
Built-in checks · regex · redaction
Know what happened.
Inspect request traces and guardrail events. Export telemetry to your observability stack.
SQLite logs · Prometheus · OTLP
02 / Universal AI gateway
Your SDK.
A different base URL.
Use OpenAI-compatible clients across native provider adapters and custom endpoints. Keep routing decisions in the gateway, not scattered through your app.
Client configurationPython
from openai import OpenAI
client = OpenAI(
base_url=GATEWAY_URL,
api_key=OMNISWITCH_KEY,
)
response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user",
"content": "Hello"}],
)- Fallback chains, retries, and timeouts
- Weighted routes and CEL conditions
- Hot-reload routes without a restart
Native provider adapters
OpenAIAnthropicGoogle GeminiGroqAzure OpenAIAWS BedrockCohere
Plus OpenAI-compatible endpoints, including Ollama and vLLM. Responses and Messages APIs are documented subsets.
03 / Governed agent infrastructure
Beyond model calls.
Into agent workflows.
Bring tool access and agent messages into a policy-aware gateway. Start with the MCP and A2A capabilities available in beta.
Govern MCP tool access
BetaFederate tool discovery and gate tool calls with policy across HTTP, SSE, and persistent stdio targets.
tools/list → policy → tools/call
Governed A2A model access
BetaSendMessage, SendMessageStream (SSE task events), GetTask, best-effort CancelTask; tasks in memory; no push notifications; no downstream agent proxy.
Agent Card → message → model pipeline
Public betaClear about what’s here.
Read the roadmapCore capabilities
Chat completions, routing, API keys, budgets, local guardrails, and request tracing.
Still evolving
MCP, A2A, OIDC/CEL authorization, and the Responses and Messages API subsets.
Know the boundaries
Single-replica SQLite control plane. Redis shares rate limits, not control state. No dashboard SSO, mTLS, SAML, or SCIM.
Make your next AI request yours.
Run OmniSwitch on your infrastructure. Start with a governed request.