Skip to content

HTTP Request Tool

Gives an AI agent the ability to send HTTP requests. Every call the model makes is gated: you see the exact request and approve or reject it before it is sent.

Exposes an HTTP client as a tool the model can call. The model decides the method, URL, headers, and body; the tool describes that request to you as a plain sentence (for example, “send a POST request to api.example.com”), and the request is only sent after you approve it. Rejected calls are never sent — the model is told you declined and continues without that action.

This tool exists so an agent can call APIs safely. Because page content an agent reads is untrusted input, a model that can both read a page and make network calls is a potential action primitive for a hostile page. The approval gate is what makes that safe. See Action Approval for the reasoning.

  • You want an agent to call an external API as part of its reasoning — fetch a record, post an update, trigger a webhook.
  • You want a human in the loop on every outbound request the model decides to make.
  • Prefer this over hand-wiring a fixed HTTP Request action node when the model should choose whether and how to call the API.
SettingNotes
AuthenticationOptional credential applied to the requests this tool makes. Resolved from the credential store at call time; the secret never travels in the workflow or the approval prompt.

The model supplies the rest of each request at call time: method (GET, POST, PUT, DELETE, PATCH, OPTIONS, HEAD), URL, optional headers, and an optional body.

Returns a tool dependency that agent nodes (such as Tools Agent) can call. It has no output port of its own — connect it into an agent’s tools input.

  • Every call is gated. The request is sent only after you approve it; there is no setting to turn the gate off.
  • The prompt names the action, not the payload. You see the method and host — never headers, bodies, or credential values.
  • Blocked and unsafe URLs are refused. The tool validates the URL before prompting or sending, and rejects internal/loopback and cloud-metadata addresses (the same SSRF protection as the built-in HTTP Request node), on the initial URL and on every redirect.
  • Unattended runs auto-reject. If no one responds within the approval timeout, the call is rejected so a background run can’t act without a human.
  • Optional Authentication credential (Basic Auth, Bearer token, etc.). Reuses the same credential types as the built-in HTTP Request node.

Connect HTTP Request Tool into the tools input of a Tools Agent. Give the agent a goal that needs an API call (“look up the order status for #1234”). When the agent decides to make the request, the run surfaces an approval row in the Executions pane naming the method and host; approve it to let the call proceed.

  • If a call never fires, check the Executions pane — it may be waiting on your approval, or it may have auto-rejected after the timeout.
  • If the model reports the action was denied, either you (or the timeout) rejected it, or the URL was blocked by the SSRF guard.
  • Verify the optional Authentication credential is set if the target API requires it.