gemini-agent-sdk-setup
通过OpenAI兼容层使用官方Google Gemini API,将OpenAI Agent SDK与Google Gemini后端集成到FastAPI服务中。
📖 中文导读:通过OpenAI兼容层使用官方Google Gemini API,将OpenAI Agent SDK与Google Gemini后端集成到FastAPI服务中。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。
name: Gemini Agent SDK Setup description: Integrate OpenAI Agent SDK with Google Gemini backend into FastAPI service using the official Google Gemini API via OpenAI compatibility layer.
Gemini Agent SDK Setup
Instructions
Generate the Agent SDK client initialization in app/services/agent/client.py:
- Import OpenAI from openai package
- Configure client with GEMINI_API_KEY
- Set base_url to Google's Gemini OpenAI compatibility endpoint: "https://generativelanguage.googleapis.com/v1beta/openai/"
- Ensure proper error handling and retry logic
Create agent configuration in app/services/agent/config.py:
- Define model name constants (gemini-1.5-flash, gemini-2.5-flash)
- Set up default parameters for agent calls
- Include proper timeout and retry configurations
Implement tool registration system in app/services/tools/registry.py:
- Create function to register tools with the agent
- Define tool schemas following OpenAI Agent SDK specification
- Include validation for tool inputs and outputs
Create utility functions for calling Gemini models in app/services/agent/utils.py:
- Wrapper function for agent completion calls
- Error handling and logging utilities
- Response parsing and validation
Follow Context7 MCP conventions:
- Only use Gemini models (no OpenAI inference)
- Follow deterministic output formatting
- Include proper error handling and logging
Examples
Input: "Setup Gemini Agent SDK integration" Output: Creates client.py with:
from openai import OpenAI
import os
client = OpenAI(
api_key=os.getenv("GEMINI_API_KEY"),
base_url="https://generativelanguage.googleapis.com/v1beta/openai/"
)
def call_gemini_agent(messages, tools=None, model="gemini-1.5-flash"):
"""Call the Gemini agent with provided messages and tools."""
try:
response = client.chat.completions.create(
model=model,
messages=messages,
tools=tools,
temperature=0.1
)
return response
except Exception as e:
# Handle errors appropriately
raise e