Native WebSearch — GPT-OSS

Some Clovis models natively integrate WebSearch capabilities. WebSearch is then driven directly by the model, which greatly simplifies integration.

Info

This approach is an alternative to MCP WebSearch. For fine-grained control over orchestration (autonomous agents, multi-tool workflows), prefer MCP WebSearch.

---

Compatible models

Model	Native WebSearch
ClovisLLM/gpt-oss-120b	✔️

---

How it works

With these models, WebSearch is natively integrated into the model's reasoning:

1. The model detects that a web search is relevant
2. It automatically performs the search
3. It leverages the results in its generation

There is no need to:

• declare an MCP server in the request,
• manage complex client-side orchestration,
• implement specific routing logic.

---

Example

Request with GPT-OSS

{
"model": "ClovisLLM/gpt-oss-120b",
"input": "What are the latest news about AI in France?",
"temperature": 0.3
}

No tools declaration is required. The model decides on its own when to perform a web search and integrates it into its response.

---

MCP WebSearch vs Native WebSearch

	MCP WebSearch	Native GPT-OSS WebSearch
Principle	The model calls an external MCP server	WebSearch is natively integrated into the model
Control	Full control over orchestration	Greatly simplifies integration
Use case	Autonomous agents, workflows, multi-tool orchestration	Quick integration without client-side orchestration
Models	All Clovis models	ClovisLLM/gpt-oss-120b, ClovisLLM/gpt-oss-200b
Declaration	Yes (tools with type: "mcp")	No (no configuration required)

Attention

Native GPT-OSS WebSearch is a specific use case. For fine-grained control over orchestration, prefer MCP WebSearch.

---

Declare the MCP server

To enable WebSearch, add an mcp type tool in the tools field of your request:

MCP declaration

{
"type": "mcp",
"server_label": "web_search_preview",
"server_description": "The websearch mcp",
"server_url": "https://llm-gateway.clovis-ai.fr/api/v1/mcp",
"require_approval": "never"
}

Field	Type	Required	Description
type	string	✔️	Must be "mcp"
server_label	string	✔️	Logical name of the MCP server
server_description	string		Functional description of the server
server_url	string	✔️	Clovis MCP server URL
require_approval	string		Execution policy (never = automatic)

---

Full example

Complete request

{
"model": "ClovisLLM",
"input": [
  {
    "role": "system",
    "content": [
      {
        "type": "input_text",
        "text": "You are a helpful assistant. If a question requires up-to-date information, use the websearch."
      }
    ]
  },
  {
    "role": "user",
    "content": [
      {
        "type": "input_text",
        "text": "What are the latest news about AI in France?"
      }
    ]
  }
],
"tools": [
  {
    "type": "mcp",
    "server_label": "web_search_preview",
    "server_description": "The websearch mcp",
    "server_url": "https://llm-gateway.clovis-ai.fr/api/v1/mcp",
    "require_approval": "never"
  }
],
"temperature": 0.3,
"max_output_tokens": 1200
}

Response

The model returns a message type response containing the synthesis of web results:

Response

{
"id": "resp_xxx",
"object": "response",
"output": [
  {
    "type": "message",
    "role": "assistant",
    "content": [
      {
        "type": "output_text",
        "text": "Here are the latest news about AI in France: ..."
      }
    ]
  }
],
"usage": {
  "input_tokens": 350,
  "output_tokens": 420,
  "total_tokens": 770
}
}

---

Code examples

Python

Python

from openai import OpenAI

CLOVIS_API_KEY = "sk-xxxx"
CLOVIS_BASE_URL = "https://llm-gateway.clovis-ai.fr/v1"
CLOVIS_MODEL_NAME = "ClovisLLM"

client = OpenAI(api_key=CLOVIS_API_KEY, base_url=CLOVIS_BASE_URL)

answer = client.responses.create(
  model=CLOVIS_MODEL_NAME,
  input="What are the latest news about AI?",
  tools=[
      {
          "type": "mcp",
          "server_label": "web_search_preview",
          "server_description": "The websearch mcp",
          "server_url": "https://llm-gateway.clovis-ai.fr/api/v1/mcp",
          "require_approval": "never",
      }
  ]
)

print(answer.output_text)

JavaScript

JavaScript

const axios = require('axios');

const CLOVIS_API_KEY = 'sk-xxxx';
const CLOVIS_BASE_URL = 'https://llm-gateway.clovis-ai.fr/v1';

async function searchWeb() {
const res = await axios.post(
  CLOVIS_BASE_URL + '/responses',
  {
    model: 'ClovisLLM',
    input: 'What are the latest news about AI?',
    tools: [
      {
        type: 'mcp',
        server_label: 'web_search_preview',
        server_url: 'https://llm-gateway.clovis-ai.fr/api/v1/mcp',
        require_approval: 'never',
      }
    ],
  },
  {
    headers: {
      Authorization: 'Bearer ' + CLOVIS_API_KEY,
      'Content-Type': 'application/json',
    },
  }
);

console.log(res.data.output[0].content[0].text);
}

searchWeb();

---