One high-performance gateway for service, LLM, and MCP traffic.
An open source HTTP and gRPC gateway that handles traditional application traffic and AI-native protocols in one data plane. Route, secure, observe, and govern services, LLM provider traffic, MCP tools, and agent-to-agent communication without stitching together separate gateways.
curl -sL https://agentgateway.dev/install | bash
LLM Gateway
Route to OpenAI, Claude, Gemini, or any self-hosted model with credentialing and failover.
MCP / A2A
Native protocol support for Model Context Protocol and agent-to-agent traffic.
APIs & Services
HTTP, gRPC, and TCP traffic — observable by default with mTLS and OIDC built in.
Swipe to explore →
Everything you need for agent traffic — in one binary. #
One control plane for service traffic, LLM routing, MCP, A2A, and inference — with the policy, observability, and cost controls that a platform team actually trusts.
ROUTE
One data plane for HTTP, gRPC, and TCP services, every major LLM provider, self-hosted inference, and the MCP and A2A protocols your agents speak.
Run any north–south or east–west traffic through a single binary. Zero-config TLS, OIDC, mTLS-rotation, native Envoy compatibility for those who want it.
Drop-in OpenAI-compatible API in front of OpenAI, Anthropic, Bedrock, Gemini, Vertex, and OSS Llama runs — with token budgets, semantic caching, and prompt redaction.
Latency-aware, cost-aware, model-aware routing for self-hosted inference. Plug-in vLLM, TGI, Triton; agentgateway picks the warmest replica.
Treat MCP servers like microservices: discover them, sign and scope every tool call, and keep an audit trail of what agents invoked.
Route agent-to-agent invocations between LangChain, CrewAI, ADK, and your own runtime, with identity and tracing on every hop.
SECURE
Authenticate, authorize, and redact on every hop — not just at the edge.
The policy layer itself: what attaches where, how rules merge, and how a decision is reached on every hop — with tamper-evident audit logs.
Validate JWTs at the edge and pull claims into routing and authorization decisions, or hand off a full OIDC login flow for browser traffic.
Issue and rotate API keys per consumer, or fall back to basic auth for internal tooling that cannot speak OIDC.
Delegate the allow/deny call to your own service over gRPC or HTTP, so existing entitlement logic stays where it already lives.
Express allow/deny rules over method, path, headers, and JWT claims — evaluated per hop, on both HTTP and network listeners.
Authenticate callers into MCP servers and scope which tools each identity may invoke, so an agent cannot reach past its mandate.
Terminate TLS at the listener and originate verified TLS or mTLS to every backend, with certificate rotation handled for you.
Browser-facing protections for gateways that serve web apps directly: origin allow-lists and CSRF token enforcement.
OBSERVE
Know exactly what happened on every request, and debug it without a restart.
OpenTelemetry by default. Token-usage histograms on Prometheus, traces in Jaeger, and request logs with status, latency, and token counts — each carrying the realized USD cost.
See what a proxy actually loaded, without a restart. proxy trace taps live requests into an interactive terminal view, proxy config dumps runtime state, and log levels change on the fly.
GOVERN
Decide what each caller may invoke, inspect what flows through, and know how overlapping rules resolve.
Decide which tools on which MCP targets a caller may reach, and rewrite or restrict what the server exposes downstream.
Inspect tool calls and responses in flight, blocking prompt injection and unsafe payloads before they reach an agent or a server.
Pattern and model-based checks on LLM prompts and completions, with redaction for anything that should never leave your estate.
COST
Attribute, price, and cap what your agents spend.
Spend control in three stages: attribute it, price it, then enforce it — before the invoice is the first place you find out.
Hard caps per key or team: cut off token consumption and dollar spend once a limit is hit, rather than alerting after the fact.
Issue scoped keys per team, app, or user so every request carries an owner — the unit that budgets and reports attach to.
Spend, token usage, and traffic in the admin UI, grouped by model, provider, and user, so you can see where the money went.
Per-model input and output pricing, so each request carries a realized USD cost that logs, metrics, traces, and policies can all read.
Getting started #
Quick start
Get started with agentgateway in less than 5 minutes — local-first.
MCP connectivity guide
Connect agentgateway to MCP servers with built-in discovery and auth.
Deployment options
Binary, Docker, or Kubernetes — pick what fits your stack.
Hands-on lab: LLM cost management
Free, on-demand lab — monitor LLM usage, apply consumption controls, and optimize costs with agentgateway.
Popular integrations #
What could be improved?
Your feedback helps us improve assistant answers and identify docs gaps we should fix.
Need more help? Join us on Discord:
https://discord.gg/y9efgEmppm Want to use your own agent? Add the Solo MCP server to query our docs directly. Get started here: https://search.solo.io/.