docs: convert the goose example from an E2E to an example (#1356)

**Description**

The goose example has been non-stop flakey tests because the model we
can use in CI is not strong enough to be predictable. This converts it
to an example instead,

Signed-off-by: Adrian Cole <[email protected]>
Co-authored-by: Takeshi Yoneda <[email protected]>

Author: codefromthecrypt

examples/README.md

@@ -17,6 +17,12 @@ Examples demonstrating how to configure the MCP Gateway to connect AI agents to
 - OAuth authentication with Keycloak
 - Combining LLM routes and MCP routes on the same Gateway
 
+### [Goose Integration](./goose/)
+
+Shows how to integrate the Goose AI agent framework with MCP tools,
+demonstrating unified routing of both LLM and MCP calls through the Envoy AI
+Gateway for single-agent origin.
+
 ## Advanced Features
 
 ### [Provider Fallback](./provider_fallback/)

examples/goose/README.md

@@ -0,0 +1,75 @@
+# Goose MCP Recipe Example
+
+## Overview
+
+This example demonstrates how to use [Goose][goose], an AI agent framework by
+Block, with the Envoy AI Gateway (aigw). It shows how to create a single agent
+that handles both LLM (Large Language Model) calls and MCP (Model Context
+Protocol) tool calls through the same gateway.
+
+Goose is an AI agent framework for tools and LLMs. MCP connects AI models to
+external tools. The [example recipe](kiwi_recipe.yaml) searches for flights
+from New York to Los Angeles on a specified date using the Kiwi flight search.
+
+This demonstrates single-agent origin, where all AI interactions, LLM and MCP,
+flow through the Envoy AI Gateway for observability, security, and routing.
+
+## Prerequisites
+
+- Ollama installed and running.
+- Goose installed.
+- aigw binary installed.
+
+## Manual Steps
+
+1. **Start Ollama** on your host machine:
+
+   Run Ollama on all interfaces with a large context size to support access from Docker and handle complex tasks.
+
+   ```bash
+   OLLAMA_CONTEXT_LENGTH=131072 OLLAMA_HOST=0.0.0.0 ollama serve
+   ```
+
+2. **Start aigw with MCP configuration**:
+
+   Launch aigw in standalone mode to route LLM requests to Ollama and MCP requests to the Kiwi flight search server.
+
+   ```bash
+   OPENAI_BASE_URL=http://localhost:11434/v1 OPENAI_API_KEY=unused aigw run --mcp-json '{"mcpServers":{"kiwi":{"type":"http","url":"https://mcp.kiwi.com"}}}'
+   ```
+
+3. **Run the Goose recipe**:
+
+   Execute the flight search recipe with Goose, directing it to the Envoy AI Gateway (port 1975) for both LLM and MCP access.
+
+   ```bash
+   OPENAI_HOST=http://127.0.0.1:1975 OPENAI_API_KEY=test-key \
+     goose run --provider openai --model qwen3:1.7b --recipe kiwi_recipe.yaml --params flight_date=31/12/2025
+   ```
+
+   - Replace `qwen3:1.7b` with your preferred Ollama model.
+   - Set `flight_date` to a future date in DD/MM/YYYY format.
+
+4. **Verify the output**:
+
+   The recipe outputs a JSON structure with the top 3 flight options from New
+   York to Los Angeles, including airline, flight number, and price.
+
+   Example output:
+
+   ```json
+   {
+     "contents": [
+       {
+         "airline": "Example Airlines",
+         "flight_number": "EA123",
+         "price": "$299"
+       },
+       ...
+     ]
+   }
+   ```
+
+---
+
+[goose]: https://block.github.io/goose/

tests/e2e-aigw/testdata/kiwi_recipe.yaml renamed to examples/goose/kiwi_recipe.yaml

@@ -31,18 +31,6 @@ prompt: |
 
   ***********ANY FLIGHT IS FINE. DO NOT CONSIDER ANY USER PREFERENCES.************
 
-# Start aigw like this to use Ollama and Kiwi MCP backends:
-# OPENAI_BASE_URL=http://localhost:11434/v1 OPENAI_API_KEY=unused aigw run --mcp-json '{
-#   "mcpServers": {
-#     "kiwi": {
-#       "type": "http",
-#       "url": "https://mcp.kiwi.com"
-#     }
-#   }
-# }'
-#
-# Then, run goose like this to use Envoy AI Gateway for both LLM and MCP:
-# OPENAI_HOST=http://127.0.0.1:1975 OPENAI_API_KEY=test-key goose run --provider openai --model qwen3:1.7b --recipe kiwi_recipe.yaml --params flight_date=31/12/2025
 extensions:
   - name: mcp_gateway
     type: streamable_http