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 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