STOPHY_API_KEY. See the Quickstart.
1. Find pages
data.web holds results with a url, title, description and position. These are leads, not evidence. See Search for sources and filters.
2. Read one page
maxChars keeps a long page from filling the agent’s context.
data.markdown. When it was cut, data.nextOffset says where the next part starts, and data.warning says so. Send the same request with "offset": <nextOffset> to read on. See Content filtering.
3. Wrap both as tools
Return text the model can read, and turn an error into text so it can try another input.Source: <url> at the top of each page is what lets the agent cite.
4. Describe the tools to the model
Registersearch_web with a required string query, and say it returns titles, URLs and snippets without reading pages. Register read_page with a required string url and an optional integer offset. Use your provider’s schema, as in LLM SDKs.
Add one rule to the system prompt: answer only from text the tools returned, and end each claim with the Source: URL it came from.
Check it
- A non-public URL such as
http://localhost/adminfails withinvalidRequest, and the loop keeps running. - The agent calls
search_webfirst, thenread_pageon one or two results. - Every URL in the answer appeared in a tool result. Open two and check the claim.
- A long page is read in parts with
offset, not guessed from the first 1,500 characters.