agentrouter

One OpenAI-compatible API for every model, one router for every MCP tool. Formerly Envoy AI Gateway.

[agentrouter @agentrouterai]

$ROUTER is now live on Pons.

Integrate once. Switch models without rewriting. Providers, credentials, limits and failover move out of your agent and into one router.

Models. Tools. Agents. One router.

CA: 0x83890309681e005506ef32887406e9c8cd9dc5f3

[agentrouter @agentrouterai]

How it works: you describe providers, credentials, tools and limits once. Agent Router turns that into Envoy configuration.

Agent Router configures. Envoy handles the traffic, the same proxy that already carries much of the world's production traffic.

[agentrouter @agentrouterai]

What is Agent Router, and why does your agent need one?

A thread on the problem, how it works, and what it does for you 🧵

[agentrouter @agentrouterai]

Agent Router moves all of that out of your agent and into configuration.

Your agent gets one OpenAI-compatible API for every model and one router for every MCP tool. Platform teams handle providers, credentials, limits and failover.

[agentrouter @agentrouterai]

The problem: every agent ends up carrying its own provider APIs, credentials, failover logic, tool wiring and token limits.

That's a lot of infrastructure living inside code that should be about the agent itself.

[agentrouter @agentrouterai]

Agent Router is built in the open: an Agentic AI Foundation project with maintainers from four companies and weekly meetings on Mondays.

1.2 is the first release under the new name. Bring the production problem you're actually running into.

[agentrouter @agentrouterai]

Your agent never counts a token.

[agentrouter @agentrouterai]

Agent Router configures: providers, credentials, routing, tools and limits. Envoy handles the traffic, as it has in production for a decade.

[agentrouter @agentrouterai]

Agent Router is open source. Read the code, run it on your laptop, ship the same config to Kubernetes.

https://t.co/sBKMBIIVCE

[agentrouter @agentrouterai]

From laptop to production, only the packaging changes.

OPENAI_API_KEY=sk-... aigw run
export OPENAI_BASE_URL=http://localhost:1975/v1

The same config ships to a gateway or a Kubernetes cluster. Or skip the install and use a hosted service.

[agentrouter @agentrouterai]

$ROUTER supply burned. 🔥

Burned for good. Back to building.

https://t.co/eeoJXsk9tA

[agentrouter @agentrouterai]

7/7 · Observability. Which model answered, how many tokens, time to first token, where a fallback kicked in. No instrumentation in your agent, OpenTelemetry GenAI conventions.

[agentrouter @agentrouterai]

One integration point for your agent. Six things that move out of your agent code and into configuration: 1/7

[agentrouter @agentrouterai]

6/7 · Inference-aware routing. Route to self-managed models with Kubernetes InferencePool support and inference-aware endpoint picking.

[agentrouter @agentrouterai]

5/7 · One router for every MCP tool. One tool catalog from many MCP servers, filtered by who is asking, so context stays small and wrong-tool calls stay rare.

[agentrouter @agentrouterai]

4/7 · Credentials stay out of your code. Your agent authenticates to the router and nothing else. Provider keys and AWS, GCP and Azure credentials are rotated in one place.

[agentrouter @agentrouterai]

3/7 · Traffic management. Provider fallback, model name virtualization, and token limits per team, app, or model. Your agent never counts a token.

[agentrouter @agentrouterai]

2/7 · One API, every provider. Route OpenAI-compatible requests to Anthropic, Bedrock, Vertex, Azure, or self-hosted vLLM. Switching models is changing the model name.

[agentrouter @agentrouterai]

Formerly Envoy AI Gateway. Same code, same maintainers.

[agentrouter @agentrouterai]

Integrate once. Switch models without rewriting.

Agent Router gives your agent one OpenAI-compatible API for every model and one router for every MCP tool. Platform teams set up providers, credentials, limits and failover. Envoy handles the traffic.