OpenAI SDK: Built-in tools

Function tools (part 07) run in your code. Built-in (hosted) tools run on OpenAI’s side within the same request: web search, file search over your uploaded documents, a code interpreter sandbox, image generation, and remote MCP servers. You enable them with one line, and the response contains items showing what the tool did. They save a lot of plumbing, at the cost of control and extra tool fees.

1. The catalogue

Type names come from the installed SDK’s type definitions (openai 2.44). Availability and pricing vary by model, so check the tools guide for yours.

2. Web search with citations

real API

from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    model="gpt-5.6-luna",
    tools=[{"type": "web_search"}],
    input="What changed in the latest Python release? Cite sources.",
)
print(response.output_text)

for item in response.output:
    if item.type == "web_search_call":
        print("searched:", item.status)
    elif item.type == "message":
        for part in item.content:
            for ann in getattr(part, "annotations", []):
                if ann.type == "url_citation":
                    print("source:", ann.title, ann.url)

3. Code interpreter: let the model compute

real API

response = client.responses.create(
    model="gpt-5.6-luna",
    tools=[{"type": "code_interpreter", "container": {"type": "auto"}}],
    instructions="Use Python for any arithmetic. Show the final numbers only.",
    input="Monthly revenue: 120, 135, 128, 160, 171, 190 (lakh). What is the average month-on-month growth rate?",
)
print(response.output_text)          # the maths was executed, not predicted token by token

Language models are poor calculators. Letting them write and run code turns arithmetic, statistics and data wrangling into something reliable. Containers are billed separately from tokens and expire after a period of inactivity.

4. Remote MCP servers

real API — a hosted MCP server's tools become available to the model

response = client.responses.create(
    model="gpt-5.6-luna",
    tools=[{
        "type": "mcp",
        "server_label": "docs",
        "server_url": "https://example.com/mcp",       # an MCP server you trust
        "allowed_tools": ["search_docs"],              # expose only what you need
        "require_approval": "always",                  # your code approves each call (safer default)
    }],
    input="How do I rotate an API key? Search our docs.",
)
for item in response.output:
    print(item.type)        # mcp_list_tools, mcp_approval_request, mcp_call, message …

With require_approval, the response contains an mcp_approval_request item. You reply with an approval item to let the call proceed. The AI Engineering guide builds an MCP server (parts 12–14) and explains why a third-party MCP server deserves the same scrutiny as any other code that can act on your behalf.

5. Hosted, function, or both?

They combine freely: one request can offer web search, file search and your own functions, and the model picks per step.

Recap