Claude tool use with web search

Anthropic's Claude models natively parse and execute client-side tools. Supplying an ODEN web search tool equips Claude with verified live internet knowledge and attributed citations, hosted within the EU.

Define the Claude tool schema#

tools = [{
    "name": "fetch_live_web_context",
    "description": "Retrieve grounded factual web information and source metadata from the live European search index.",
    "input_schema": {
        "type": "object",
        "properties": {
            "search_query": {"type": "string", "description": "The specific question or topical phrase"}
        },
        "required": ["search_query"],
    },
}]

Handle the tool_use block with XML tags#

Claude excels at parsing structured XML containers. Format the retrieved ODEN answer and citation metadata inside semantic XML tags:

import os, requests
from anthropic import Anthropic

client = Anthropic()

def search_oden_knowledge(topic: str) -> str:
    res = requests.post(
        "https://api.oden-api.com/search",
        headers={"Authorization": f"Bearer {os.environ['ODEN_KEY']}"},
        json={"query": topic, "include_snippets": True, "depth": "advanced"},
        timeout=25,
    )
    res.raise_for_status()
    payload = res.json()["results"]
    ref_list = "\n".join(
        f"[{i+1}] {c['title']} <{c['url']}>"
        for i, c in enumerate(payload["citations"])
    )
    return (
        f"<web_context>\n"
        f"<summary>{payload.get('answer', '')}</summary>\n"
        f"<citations>\n{ref_list}\n</citations>\n"
        f"</web_context>"
    )

messages = [{"role": "user", "content": "What are the latest 2026 corporate sustainability reporting deadlines in Sweden?"}]
resp = client.messages.create(model="claude-sonnet-5", max_tokens=1024, tools=tools, messages=messages)

if resp.stop_reason == "tool_use":
    tool_use = next(b for b in resp.content if b.type == "tool_use")
    result_context = search_oden_knowledge(tool_use.input["search_query"])
    messages.append({"role": "assistant", "content": resp.content})
    messages.append({"role": "user", "content": [{
        "type": "tool_result",
        "tool_use_id": tool_use.id,
        "content": result_context,
    }]})
    final = client.messages.create(model="claude-sonnet-5", max_tokens=1024, tools=tools, messages=messages)
    print("".join(b.text for b in final.content if b.type == "text"))

Claude thinking mode & reasoning token coordination#

When invoking Claude with extended thinking mode enabled, Claude can reason through complex search requirements, formulate multi-hop queries, and analyze contradictory web sources:

# Example configuring Anthropic client with thinking budget and tool use
response = client.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=4096,
    thinking={
        "type": "enabled",
        "budget_tokens": 2048
    },
    tools=tools,
    messages=[{
        "role": "user",
        "content": "Synthesize recent legal interpretations of EU AI Act open source model exemptions."
    }]
)

Forcing tool execution with tool_choice#

If your workflow requires guaranteed factual retrieval before Claude generates an answer, enforce tool execution using the tool_choice parameter:

# Enforce that Claude MUST execute the web_search tool
resp = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=1024,
    tools=tools,
    tool_choice={"type": "tool", "name": "web_search"},
    messages=messages
)

System prompt recommendations for citation quality#

Instruct Claude explicitly to format and ground responses:

system_prompt = (
    "You are a factual research assistant. When web search results are provided, "
    "synthesize the findings concisely and attribute key facts with inline brackets [1], [2]. "
    "If candidate search sources contain conflicting assertions, explicitly report the discrepancy."
)

Notes#

• Append the assistant's tool_use turn and your tool_result before the follow-up call — Claude needs both to continue.
• Ask Claude in the system prompt to cite the returned sources inline.
• ODEN answers in the query's language, which pairs well with Claude's multilingual output.

FAQ#

Does ODEN work with Claude's own web search tool?#

They are alternatives. Claude's built-in web tools are one option; ODEN is a dedicated, EU-hosted European alternative you control and meter yourself.

Can I stream Claude's final answer?#

Yes. Stream the final messages.create call as usual — ODEN's result is already in the message history by then. ODEN itself returns a complete response rather than streaming.

Which Claude models support tool use?#

The current Claude models all support tool use with the tools parameter shown above.