Agent Gateway KrakenD

A KrakenD based Agent Gateway implementation that serves as an egress API gateway for routing incoming requests to exposed agents within the agentic platform.

Plugins

Following plugins are included in this repository:

• Agent Card URL Rewriting Plugin
• OpenAI A2A Plugin

Development

Prerequisites

The following tools are required for development:

• Docker: For containerization and local development

Building Plugins

make plugins

This will compile the plugins and output it to build/.

Run with Docker Compose

Start the agent gateway using Docker Compose:

# Then start the services
docker compose up --build

Stop the services:

docker compose down

Testing

OpenAI-Compatible API

The gateway provides OpenAI-compatible endpoints (/models and /chat/completions) for agent access. For comprehensive documentation including request/response formats, configuration options, and protocol transformation details, see the OpenAI to A2A Plugin README.

Quick Examples:

List available agents:

curl http://localhost:10000/models

Send a chat completion request:

curl http://localhost:10000/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "local/mock-agent",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

For detailed API documentation, agent routing behavior, and model parameter formats, refer to the plugin documentation linked above.

A2A via JSON-RPC

Verify the gateway proxy functionality with a JSON-RPC message:

curl http://localhost:10000/local/mock-agent \
  -H "Content-Type: application/json" \
  -d '{
     "jsonrpc": "2.0",
     "id": 1,
     "method": "message/send",
     "params": {
       "message": {
         "role": "user",
         "parts": [
           {
             "kind": "text",
             "text": "Hello, mock agent!"
           }
         ],
         "messageId": "9229e770-767c-417b-a0b0-f0741243c589",
         "contextId": "abcd1234-5678-90ab-cdef-1234567890ab"
       },
       "metadata": {}
     }
   }' | jq

Migration Guide

Breaking Change: Removal of Per-Agent Chat Completions Endpoints

Effective from version: v0.5.0

The legacy per-agent chat completions endpoints have been removed in favor of the standard OpenAI-compatible global endpoint.

What Changed

Old Endpoint Pattern (Removed):

POST /{agent-name}/chat/completions

New Endpoint Pattern:

POST /chat/completions

Migration Steps

Before:

# Old endpoint - NO LONGER WORKS
curl http://localhost:10000/mock-agent/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4",
    "messages": [
      {
        "role": "user",
        "content": "Hello, how can you help me?"
      }
    ]
  }'

After:

# New endpoint - use model parameter to specify agent
curl http://localhost:10000/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "local/mock-agent",
    "messages": [
      {
        "role": "user",
        "content": "Hello, how can you help me?"
      }
    ]
  }'

Key Changes

1. Endpoint URL: Use /chat/completions instead of /{agent-name}/chat/completions
2. Model Parameter: The model field now specifies which agent to route to.
3. Standardization: The new endpoint follows the OpenAI API specification exactly

Benefits

• OpenAI Compatibility: Full compatibility with OpenAI client libraries and tools
• Simplified API: Single endpoint for all agents reduces API surface

Contribution

See Contribution Guide for details on contribution, and the process for submitting pull requests.