Envoy AI Gateway (now Agent Router)

Renamed

The project README (read 2026-10-08) says Envoy AI Gateway is now Agent Router, an Agentic AI Foundation project: same code, maintainers, release cadence and Apache 2.0 licence. The repo moved to theagentrouter/agent-router (old envoyproxy/ai-gateway links redirect), the site to theagentrouter.ai. CRDs, API group (aigateway.envoyproxy.io), the aigw CLI, images and Helm charts keep their names. The rename date is not stated in the pages read.

What it is

An open-source control plane for AI and agent traffic built on Envoy and Envoy Gateway: one OpenAI-compatible API for hosted providers, self-hosted inference and MCP servers, with credentials, routing, quotas, failover and usage attribution in one place.

Maker and licence

Apache-2.0. Repo theagentrouter/agent-router (formerly envoyproxy/ai-gateway): 2,200 stars on 2026-10-08, latest release v1.2.0 (2026-10-07) per GitHub API. Governance: Agentic AI Foundation (see Agentic AI Foundation). The foundation’s blog post of 2026-09-09 announces the rename and the move: “Envoy AI Gateway is now Agent Router and is joining the Agentic AI Foundation”. Per that post nine maintainer seats are shared across Bloomberg, Nutanix, AMD, Tetrate and Netflix with no company holding a majority, the project has 132 contributors from 21 organisations, Bloomberg and Tetrate began the work in 2024, v1.0 shipped in June 2026 with a compatibility commitment for 1.x, and v1.1 followed in August. aaif.io/projects lists Agent Router as a hosted project alongside agentgateway.

Position

LLM gateway and MCP gateway, Kubernetes-native (Gateway API / Envoy Gateway), plus a standalone aigw run CLI mode.

MCP gateway (docs, read 2026-10-08)

  • MCPRoute API aggregates several MCP servers behind one endpoint; tool names are prefixed with the backend name for routing.
  • Security: OAuth authentication, authorization with JWT claims, scopes and CEL expressions, upstream API-key / header injection, tool filtering.
  • Transport: streamable HTTP; the docs claim compliance with the June 2025 MCP spec including prompts, resources, notifications and server-to-client requests (vendor statement).
  • Observability: OpenTelemetry tracing and Prometheus metrics; Envoy load balancing, rate limiting, circuit breaking.
  • LLM side: many providers (OpenAI, Azure OpenAI, Gemini, Vertex AI, Bedrock, Anthropic, Mistral and others listed in the README), two-tier gateway pattern for self-hosted inference.

Maturity

Release v1.2.0 on 2026-10-07 (GitHub API); docs version 1.2. Whether the MCP features carry a GA label is not stated in the pages read.

Fit and limits (opinion)

Fits teams already running Envoy/Envoy Gateway on Kubernetes and wanting model and tool traffic under the same data plane. It assumes Kubernetes skills; for a laptop or single-host setup lighter gateways are simpler. Compare agentgateway, another gateway under the same foundation.

Self-learning

Sources

Open items

  • The “CNCF” association with the original Envoy AI Gateway was not re-verified and is not claimed; the foundation post names the Agentic AI Foundation as the home.
  • Newer MCP spec revisions beyond June 2025: not stated.