agentgateway Documentation — Open-Source Gateway for MCP, A2A and LLM Traffic

agentgateway is a Linux Foundation project written in Rust that gives agent traffic a single policy and observability point, and its documentation is published as two parallel trees: a standalone binary or Docker deployment for local work and simple production setups, and a Kubernetes deployment driven by a built-in control plane and the Kubernetes Gateway API, versioned through 1.5.x. Both trees follow the same shape. A get-started track installs the proxy and puts it in front of an OpenAI-compatible LLM, an MCP server, and ordinary non-agentic HTTP traffic. The About and Policies sections explain the architecture, how policies attach to listeners and routes, policy processing order, targeting and merging, and conditional policies. The LLM consumption track is the largest: model aliasing and virtual models, provider and API-key management, load balancing and model failover, content-based routing, streaming and OpenAI Realtime, function calling, prompt enrichment and prompt templates, request transformations, rate limiting, CEL-based RBAC, cost controls and guardrails, plus API-type pages for chat completions, responses, embeddings, rerank, token counting and Gemini. The MCP connectivity track covers static, dynamic and virtual MCP servers, MCP Apps, per-tool access control, stateful MCP, JWT auth for services, MCP guardrails and spec compatibility; the agent connectivity track covers connecting to an agent and to AWS Bedrock AgentCore. Remaining sections cover traffic management (matching, rewrites, redirects, traffic splitting, route delegation), resiliency (retries, timeouts, backend health, fault injection, keepalive), security (external auth, API key auth, CORS, CSRF, backend TLS, local and global rate limiting), an OpenTelemetry observability stack with metrics, traces and access logs, the agctl CLI for debugging and request tracing, and a CEL expression and Helm value reference. The repository carries roughly 4,800 stars and is actively developed with public community meetings and a published roadmap.

Topic

Frameworks

Level

Advanced

Format

Documentation

Cost

Free

Duration

~3-5 hours for the get-started, LLM and MCP tracks; reference thereafter

Provider

agentgateway (Linux Foundation)

Hands-on

Yes — code/exercises

Certificate

None

• ✓Put a single proxy in front of LLM providers, MCP servers and agent-to-agent traffic
• ✓Configure model aliasing, load balancing and automatic failover across LLM providers
• ✓Apply CEL-based RBAC and JWT or OAuth authentication to individual MCP tool calls
• ✓Enforce cost controls and rate limits before a request ever reaches a paid provider
• ✓Expose static, dynamic and virtual MCP servers to clients from one endpoint
• ✓Emit OpenTelemetry metrics, traces and access logs for every hop an agent makes
• ✓Deploy on Kubernetes through the Gateway API using Helm, ArgoCD or FluxCD
• ✓Debug a live configuration and trace individual requests with the agctl CLI

• •A Linux Foundation project rather than a single-vendor product — governance, roadmap and community meetings are public
• •Rust data plane built on the Kubernetes Gateway API, so it behaves like real ingress instead of a Python sidecar
• •Covers all three agent protocols in one place: OpenAI-compatible LLM traffic, MCP, and A2A
• •Guardrails, RBAC and cost controls are enforced in the proxy, so they apply to every framework and language behind it
• •Two complete documentation trees, so you are not reading Kubernetes docs in order to run a local binary
• •Roughly 4,800 GitHub stars, versioned docs through 1.5.x and active development

This page was written from 3 sources, 1 on domains other than agentgateway.dev.

1. 1.agentgateway.dev — docsvendor
2. 2.agentgateway.dev — llms.txtvendor
3. 3.github.com — agentgateway