Web search: Answers from the sites you trust

The demo in this video was recorded in September 2026. It ran on Claude Sonnet 5 (claude-sonnet-5). Results may differ on newer models.

Claude answers from what it learned in training, so it doesn't know about anything newer than that, or anything that was never in it. When a question needs those facts, someone has to look them up, and that legwork falls to you.

With the web search tool, Claude does that legwork itself, searching the web while it works on your request and answering from what it finds.

One of Meridian's proposals cites four Canadian engineering standards by clause, and Meridian carries professional liability if any reference is wrong. So before the proposal goes out, Meridian wants each reference checked against official sources.

Add the tool to your request

Web search is a server tool. It runs on Anthropic's infrastructure, so you don't write a handler for it or send back a tool_result.

Adding it takes one entry in your tools list. The entry's type picks the version of the tool, and Meridian sets it to web_search_20260209. Its name is always web_search. The tool reference lists the current versions, so copy the type value from there.

On Claude 4.6 and later models, web_search_20260209 and later versions use dynamic filtering by default. Claude runs code that filters the search results, so only relevant content reaches the context window.

Claude usually does all its searching during one request. Your code sends that request and gets back the finished answer.

The API can pause a long-running turn and return a stop_reason of pause_turn. When it does, send the response's content back as an assistant message, with the same tools, and Claude continues.

Each search is billed on top of what you pay for tokens. Not every cloud platform offers web search.

Name the sites Claude may use

So far, nothing in the request limits which sites Claude can search, and that's a problem for Meridian. The open web is full of forum posts and reseller sites, but Meridian can stand behind only official sources.

So you tell the search where it may look. allowed_domains is another field in the tool's entry, and it limits search results to the sites you name. Meridian names four sites. Two belong to the bodies that publish these standards and certify to them, csagroup.org and cwbgroup.org. Two belong to governments, ontario.ca and canada.ca.

allowed_domains takes a list. Write each item as a bare domain with no https://, and its subdomains are included.

Read the answer and its sources

Claude's answer can arrive in several text blocks, with the searches in between. So join every text block to get the whole answer.

Each search also adds a web_search_tool_result block to the response. It holds the list of results the search returned, and each result has a title and a url.

Of Meridian's four references, Claude found that two check out, a third checks out except for its subclause number, and the fourth is wrong.

Put the program together

Meridian's program puts these steps together. In the program, excerpt holds the passage from the proposal, and QUESTION holds Meridian's question.

Key takeaways

• Web search is a server tool, so you add one entry to tools and don't write a handler. Claude usually does all its searching during one request.
• allowed_domains limits search results to the sites you name.
• On Claude 4.6 and later models, dynamic filtering is on by default in web_search_20260209 and later versions, so only relevant content reaches the context window.