agentgateway (US, an Agentic AI Foundation project) is a data-plane gateway written in Rust, shipped as a single binary or on Kubernetes. It brings an OpenAI-compatible LLM gateway, an MCP tool gateway and agent-to-agent (A2A) support into the same data plane.
What it does
- Exposes an OpenAI-compatible endpoint in front of different providers
- Picks models by cost, latency or availability, with automatic failover
- Sets budgets and limits per key, team or period
- Applies prompt, PII and moderation guardrails before calling the provider
- Centralizes the MCP tool catalog with per-agent scoping
- Exports metrics and traces through OpenTelemetry
How it works
- The gateway runs next to the application, with no vendor-cloud middle layer
- Providers, models and policies are declared once and applied per request
- Virtual keys isolate access per project or team
- The same core handles LLM calls, tool calls and agent-to-agent traffic
- The project was donated to the Agentic AI Foundation and keeps the Apache-2.0 license
Models and providers
- First-class providers such as OpenAI, Anthropic, Bedrock, Gemini, Vertex AI, Groq, Mistral, xAI, Together and Fireworks
- Self-hosted endpoints such as vLLM and TGI join as an extra provider
- OpenAI-compatible providers can be added through configuration
Availability and license
- Open source under Apache-2.0, with no license fee
- Distributed as a standalone binary, container or Kubernetes deployment
- Solo.io maintains an enterprise distribution with support
Points to note
- Running it requires familiarity with Kubernetes or gateway configuration
- Enterprise governance features live in the commercial distribution
- It should not be confused with Agent Router, another AAIF project