Overview
Agent Router is an open-source control plane for agentic traffic, formerly known as Envoy AI Gateway. An agent uses one API while Agent Router converts provider, credential, routing, tool, and limit settings into Envoy configuration; Envoy then handles the requests.
The project is intended to separate traffic concerns from agent code without requiring users to become Envoy experts. It supports local use, dedicated gateways, Kubernetes deployments, and a hosted option.
Key Features
- Routes OpenAI-compatible requests to providers including Anthropic, Bedrock, Vertex, Azure, and self-hosted vLLM.
- Supports 17 providers out of the box and allows model switching through the model name.
- Provides provider fallback, model-name virtualization, and token limits by team, application, or model.
- Keeps provider API keys and cloud credentials in the router rather than agent code.
- Combines tools from multiple MCP servers into a filtered catalog sized for each agent.
- Supports Kubernetes InferencePool and inference-aware endpoint selection for self-managed models.
Traffic Visibility and Control
Agent Router exposes which model answered, token usage, time to first token, and where fallback occurred without requiring instrumentation inside the agent. Its telemetry follows OpenTelemetry GenAI conventions.
Credential rotation is centralized, while the agent authenticates only to the router. This design also moves provider-specific APIs, tool wiring, and traffic policies into configuration shared across agents.
Deployment Options
The same configuration can begin on a laptop and later move to a dedicated gateway or Kubernetes cluster, with packaging as the main difference. A hosted service is available for teams that do not want to manage infrastructure.
Commercial self-managed offerings add options such as cross-team cost attribution, administrative controls, runtime guardrails, enterprise single sign-on, multi-region deployment, and operation across cloud, on-premises, or edge environments.
Open Development
Agent Router is developed publicly through weekly meetings, design proposals, and maintainers from four companies. The project announcement states that its transition from Envoy AI Gateway retained the same code, maintainers, and APIs.
