agent-gateway

One gateway. Every agent protocol. Route requests between A2A, MCP, OpenAI Assistants, and plain REST — with a single YAML config.

pip install agentic-gateway

🎬 Demo

Starting the gateway, routing requests across 4 protocols, and viewing live metrics

The Problem

The AI agent ecosystem is fragmented:

Your App ──→ Agent A (A2A protocol)
         ──→ Agent B (MCP tools)
         ──→ Agent C (OpenAI Assistants API)
         ──→ Agent D (plain REST)

Each speaks a different language. You write adapters for each. Then a new protocol drops and you rewrite everything.

The Solution

                  ┌────────────────────┐
Your App ────────→│   agent-gateway    │────→ Agent A (A2A)
                  │                    │────→ Agent B (MCP)
                  │   One protocol     │────→ Agent C (OpenAI)
                  │   in, any out      │────→ Agent D (REST)
                  └────────────────────┘

Talk to the gateway in any protocol. It translates to any other.

Quick Start

1. Define your gateway

# gateway.yaml
gateway:
  name: my-gateway
  port: 8080

routes:
  - name: search-agent
    upstream:
      protocol: a2a
      url: http://localhost:9000
    expose_as:
      - mcp      # Expose A2A agent as MCP tool
      - rest     # And as REST endpoint

  - name: code-helper
    upstream:
      protocol: mcp
      command: python -m code_helper_mcp
      transport: stdio
    expose_as:
      - a2a      # Expose MCP server as A2A agent
      - openai   # And as OpenAI Assistants-compatible

  - name: legacy-api
    upstream:
      protocol: rest
      url: https://api.example.com/analyze
      method: POST
      headers:
        Authorization: "Bearer ${API_KEY}"
    expose_as:
      - a2a
      - mcp

2. Start the gateway

agent-gateway serve gateway.yaml

3. Connect from any protocol

# As A2A client:
response = a2a_client.send_task("search-agent", "Find recent papers on RLHF")

# As MCP client (from Claude Desktop):
# Just add to claude_desktop_config.json:
#   {"mcpServers": {"gateway": {"url": "http://localhost:8080/mcp"}}}

# As OpenAI-compatible:
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8080/openai/v1")
response = client.chat.completions.create(
    model="code-helper",
    messages=[{"role": "user", "content": "Fix this bug"}],
)

# As REST:
import httpx
response = httpx.post("http://localhost:8080/rest/search-agent", json={"query": "RLHF papers"})

Protocol Support Matrix

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        agent-gateway                             │
│                                                                  │
│  ┌──────────┐    ┌────────────┐    ┌──────────┐    ┌─────────┐ │
│  │ Ingress  │───→│  Router    │───→│Translator│───→│  Egress │ │
│  │          │    │            │    │          │    │         │ │
│  │ • A2A    │    │ • Match    │    │ • A2A↔X  │    │ • HTTP  │ │
│  │ • MCP    │    │   route    │    │ • MCP↔X  │    │ • stdio │ │
│  │ • OpenAI │    │ • Auth     │    │ • OAI↔X  │    │ • SSE   │ │
│  │ • REST   │    │ • Rate     │    │ • REST↔X │    │         │ │
│  └──────────┘    └────────────┘    └──────────┘    └─────────┘ │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Features

• Protocol translation — Seamless conversion between A2A, MCP, OpenAI, REST
• YAML configuration — No code required, just declare routes
• Environment variables — ${VAR} expansion in config
• Auth passthrough — Forward headers, tokens, API keys
• Health checks — Built-in /health endpoint with upstream status
• Hot reload — Change config without restart (--watch)
• Middleware — Rate limiting, logging, request/response transforms
• Discovery — Auto-generates agent cards, tool lists, OpenAPI specs

CLI

# Start gateway
agent-gateway serve gateway.yaml

# Validate config
agent-gateway validate gateway.yaml

# List routes
agent-gateway routes gateway.yaml

# Health check
agent-gateway health http://localhost:8080

Programmatic Usage

from agent_gateway import Gateway

gateway = Gateway.from_yaml("gateway.yaml")

# Translate a single request
result = await gateway.translate(
    source_protocol="rest",
    target_route="search-agent",
    payload={"query": "test"},
)

Why Not Just Use X?

agent-gateway understands the semantics of each protocol — agent cards, tool schemas, function calling, streaming — and translates them correctly.

Contributing

git clone https://github.com/naveenkumarbaskaran/agent-gateway.git
cd agent-gateway
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest

License

MIT