WebSearch vs WebFetch
People often think WebSearch reads the pages it finds. It does not. WebSearch only brings back titles and links. To read a page, Claude has to call WebFetch on one link. WebFetch then opens that page and pulls out only the part Claude asked about. So a search finds where to look, and a fetch does the reading.
WebSearch Tool
The tool that searches the web and returns titles and links.
The WebSearch tool searches the web and returns the titles and links of the results. It does not open the pages. To read one, Claude follows up with WebFetch. One call can run several searches behind the scenes to sharpen the results. An interactive session has a limit of 200 searches. Its rule has no pattern: you allow or deny WebSearch as a whole.
WebFetch Tool
The tool that reads one web page and pulls out what Claude asked for.
The WebFetch tool opens one web page and turns it into plain text. It then runs a short question about the page through a second, smaller model, and Claude gets that answer, not the whole page. So the result only holds what the question asked for. It reads up to 100,000 characters per call and keeps each page for 15 minutes. It will not open localhost. A rule like WebFetch(domain:example.com) allows one site.
In one line
- WebSearch Tool
- The tool that searches the web and returns titles and links.
- WebFetch Tool
- The tool that reads one web page and pulls out what Claude asked for.