How the agent reads the web
Ask the chat to look something up or to read a page, and the agent picks the best way to get it. You do not choose a tool. The answer says which way the page was read.
The order the agent tries
The model's own web search and fetch. Claude searches with WebSearch and fetches with WebFetch; ChatGPT models use their own search. This is the fastest way, and it is enough for most public pages.
Your own browser or computer. If a page fails, blocks robots, or needs your login, the agent calls
web_read. Hydra then asks your connected device to read the page:- first Hydra for Chrome and Firefox, your browser, signed in as you;
- then hydra-cli or Hydra for Mac and Windows, your computer's network, without your logins.
The tool's answer starts with "Read through your computer" and the kind of device that read it, such as (hydra-ext) or (hydra-cli).
Hydra's server. When none of your devices is online,
web_readreads the page from Hydra's server. The answer starts with "Read from the Hydra server". The server can read only public pages, and some shops refuse it.
Which path reads what
| Page | Model's search | Your browser | Your computer | Hydra's server |
|---|---|---|---|---|
| News, docs, most public sites | yes | yes | yes | yes |
| A shop that blocks robots | often no | yes | no | no |
| A supplier or carrier portal you are signed in to | no | yes | no | no |
| A site that only works on your office network | no | yes | yes | no |
| Your home router, NAS or another private address | no | no | no | no |
Pages on private addresses are refused on every path, so a page cannot trick the agent into reading your local network.
What the agent can do with a page
Read only. Every device fetches with a plain GET: it never submits a form, clicks a button, buys or posts. A page that tries to steer the agent ("ignore your instructions...") can at most make it read another page, never act for you.
The one exception is a step-by-step action in your own Chrome, which always shows you a card first and runs only after you press Approve. See Hydra for Chrome and Firefox.
What is recorded
For each page read, Hydra keeps the site's host name and which kind of device read it, in the audit log. It does not keep the full address or the page.
When a page still cannot be read
- "The site answered HTTP 429" or 403: the site blocks automated readers. Connect Hydra for Chrome and Firefox and ask again; your browser reads it the way you do.
- "That link is a PDF, not a web page": put the file in Google Drive and give the chat the Drive link, or attach the file.
- A page built by JavaScript comes back nearly empty: with the extension connected, Hydra opens the page in a hidden tab of your browser and reads it once it has rendered.