Web search for n8n AI agents with the HTTP Request tool

Why the HTTP Request node rather than a custom node?

Because one GET request with a header and four query parameters is all the API needs. n8n's documentation says the HTTP Request node can run as a regular node or be attached to an AI agent as a tool, and that some options exist only in the tool form.

Nothing of ours has to be installed or updated, and the key stays in n8n's credential store instead of the workflow. Give n8n a key of its own. Make one API key per agent, app or teammate and revoke any of them from the dashboard without touching the others.

How does the model write the search query?

Through n8n's $fromAI() function in the value of the q field. Its documented form is $fromAI(key, description, type, defaultValue): use the key query, a description such as "What to search for on the web", and the type string. n8n notes that $fromAI() only works in tools connected to the AI Agent node.

Describe the tool itself in plain words too, for example "Searches the web and returns sources with their page text; cite the URLs". The agent uses that text to decide when to search.

Why turn on Optimize Response?

Because everything a tool returns lands in the model's context. Optimize Response is a tool-only option that reduces what is passed to the model. With the JSON type, set Field Containing Data to results, then include only title, url and content.

The agent then sees a short list of sources with their text, and nothing else. Cap the length on our side with max_chars, which trims each page before it leaves the API.

Can I run searches on a schedule without an agent?

Yes. Use the HTTP Request node as a regular node behind a Schedule Trigger, with a fixed query. Add topic=news and time_range=day to collect the day's articles on a subject, then send the titles and URLs to a chat, a sheet or an email.

One scheduled news search per day uses about 30 credits a month, so 5,000 searches ($9.25) covers many topics for a long time. Search credits have no expiry date: what you do not use this month is still there next year.

What does the workflow receive when a page is refused?

The search still succeeds. The refused result keeps its title and URL, has no content, and its content_status says why, such as blocked_by_robots. Our reader respects robots.txt: it introduces itself as LLMWebSearch, follows the rules a site writes for it or for all robots, and leaves out the content of any page it is not allowed to read.

HTTP errors are different: a wrong key returns 401, an empty balance 402, too many calls 429. In an agent, let the tool hand the error back so the model can explain it. In a scheduled workflow, route errors to a branch that logs them. The errors page lists every code.

How do I test the call before building the workflow?

Run the same request with curl from the curl recipes. If it works in a terminal and fails in n8n, the cause sits in the node settings. The usual suspects are a credential value without the Bearer prefix, or an empty q field.