Post

Agentgateway (Solo.io / Linux Foundation)

Agentgateway is the most complete open-source agent gateway -- it natively speaks A2A and MCP, handles all three gateway layers (LLM, MCP, agent), and is now a Linux Foundation project.

Agentgateway (Solo.io / Linux Foundation)

Agentgateway is the most complete open-source agent gateway – it natively speaks A2A and MCP, handles all three gateway layers (LLM, MCP, agent), and is now a Linux Foundation project. If you want one open-source gateway for everything, this is it.


What It Is

Agentgateway is an open-source, Rust-based proxy purpose-built for AI agent communication. Created by Solo.io and contributed to the Linux Foundation in August 2025, it is designed from the ground up for agentic systems – not an API gateway with AI plugins bolted on.

It natively supports three traffic types:

  • LLM Gateway: Route to LLM providers via OpenAI-compatible API
  • MCP Gateway: Connect agents to tools via MCP (all transports: stdio, HTTP, SSE, Streamable HTTP)
  • A2A Gateway: Agent-to-agent communication via Google’s A2A protocol

This makes it the only open-source project that covers all three gateway layers in one binary.


Key Features

A2A Gateway (Agent-to-Agent)

  • Native A2A protocol support
  • Agent capability discovery
  • Modality negotiation (text, image, audio)
  • Task collaboration and lifecycle management
  • Route between agents based on capabilities

MCP Gateway (Agent-to-Tool)

  • Tool federation across multiple MCP servers
  • All MCP transports: stdio, HTTP/SSE, Streamable HTTP
  • OpenAPI to MCP integration (expose REST APIs as MCP tools)
  • OAuth authentication for MCP servers

LLM Gateway (Agent-to-Model)

  • Unified OpenAI-compatible API to major providers (OpenAI, Anthropic, Gemini, Bedrock, etc.)
  • Budget and spend controls
  • Prompt enrichment
  • Load balancing and failover across providers

Inference Routing

Kubernetes Inference Gateway extensions for routing to self-hosted models:

  • GPU utilization-aware routing
  • KV-cache-aware endpoint selection
  • LoRA adapter routing
  • Queue depth balancing

Security

  • RBAC with fine-grained policies
  • JWT authentication
  • TLS encryption
  • CEL-based access policies (Common Expression Language) for complex rules

Observability

  • OpenTelemetry integration
  • Per-agent, per-tool, per-model metrics
  • Full request/response tracing

Architecture

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Agents / Applications
        |
        v
+------------------------------------------+
| Agentgateway (Rust binary)                |
|                                           |
|  +-------------+----------+------------+ |
|  | A2A Gateway  | MCP Gate | LLM Gateway| |
|  | agent<>agent | agent>   | agent>     | |
|  | routing      | tools    | models     | |
|  +-------------+----------+------------+ |
|                                           |
|  [Auth: JWT, RBAC, CEL policies]         |
|  [Rate Limiting]                          |
|  [Observability: OTel]                   |
+------------------------------------------+
        |              |              |
        v              v              v
   Other Agents    MCP Servers    LLM Providers

Self-Hosting & Pricing

Fully open-source (Apache 2.0). Linux Foundation project – no vendor lock-in.

Component License Cost
Agentgateway Apache 2.0 Free
All features Apache 2.0 Free

Solo.io offers Solo Enterprise for agentgateway with additional enterprise features (advanced analytics, support, SLAs) but the open-source version is fully functional.

Deployment: single Rust binary via Docker, Kubernetes, or bare metal. Minimal resource footprint.


Comparison with Other Agent Gateways

Feature Agentgateway Kong (A2A plugin) Envoy AI GW APISIX
A2A native Yes Yes (plugin) No No
MCP native Yes Yes (plugin) Yes No
LLM routing Yes Yes Yes Yes
Inference routing Yes (K8s) No Yes (EPP) No
Open source (full) Yes No (enterprise) Yes Yes
Self-hosted Yes Yes (license) Yes (K8s only) Yes
Language Rust C/Lua Go/C++ C/Lua
Maturity Early (2025) Mature Early (2025) Mature (API), early (AI)

Limitations

  • Young project – contributed to LF in August 2025, still rapidly evolving
  • Smaller community than Kong or Envoy – fewer production case studies
  • No built-in guardrails – no PII detection, content filtering (relies on external services)
  • No semantic caching – not a focus of the project
  • No admin UI / dashboard – CLI and config-driven (Solo Enterprise adds UI)

When to Use

Strong fit:

  • Want one open-source gateway for all three layers (A2A + MCP + LLM)
  • Building multi-agent systems with A2A protocol
  • Need agent-to-agent capability discovery and routing
  • Kubernetes-native with self-hosted model inference
  • Want LF-backed, vendor-neutral infrastructure

Weak fit:

  • Need production-proven, battle-tested gateway today – Kong is more mature
  • Need guardrails, semantic caching, virtual keys – use Portkey or Kong
  • Not building multi-agent systems – LiteLLM or Cloudflare is simpler for LLM-only
  • Need admin dashboard / UI out of the box

References

This post is licensed under CC BY 4.0 by the author.