Function Tool Calling: Let Your AI Agent Use APIs

Any API with a public HTTPS endpoint, using GET, POST, PUT or DELETE. You define the endpoint, the method, and what it does in plain language, and the agent decides when a conversation actually needs it — there's no list of pre-built integrations to pick from. What won't work: plain HTTP, and anything that only resolves to a private, internal or loopback address. Both are blocked on purpose, so an on-premise system needs a public, authenticated endpoint in front of it first.

Three modes are supported: no auth, a stored credential attached automatically via a header, or the signed-in visitor's own token forwarded as a Bearer header. Secrets are only ever allowed in headers, never the URL or request body.

Each call has a 12-second timeout. Only GET is retried automatically, with exponential backoff, and only on a rate limit (429), a timeout (408) or a server error (5xx). Every other method is sent exactly once — PUT and DELETE included, even though HTTP treats them as idempotent — so a slow "create booking" call can't silently double-book. A bad or expired credential fails immediately instead of retrying uselessly.

Yes. A single turn can chain multiple tool calls — for example, checking availability, then confirming a booking — so the agent can complete a full request, not just answer a question.

Testing behaves differently depending on the method. A GET tool is genuinely called against your endpoint using sample values you provide for that probe, so you see the real response. A POST, PUT or DELETE is built and shown to you as a redacted preview, but never sent. Alongside it Robofy runs a read-only GET against the same URL with your resolved auth headers, which reports one of three things: the credential was rejected (401 or 403), the credential was accepted (a success response), or the check was inconclusive — write-only endpoints often don't answer GET, and a redirect isn't followed. Either way the write never happens, and a tool stays a draft until its test passes.

Yes. Any tool can be set to ask the visitor for approval before it runs, and you write the sentence shown on that approval card. It relies on the widget's confirmation UI, so it works on the web widget and not on WhatsApp.

Up to 15 custom tools per agent. If you have several agents, you can build a tool once and export it as JSON, then import it on another agent. Credential references come across with it; the secrets themselves stay behind.

Standard tools work on both. Two kinds are web-widget only: tools that require per-call visitor approval, because WhatsApp has no confirmation card, and tools authenticated as the signed-in visitor, because that identity comes from your web page. Those are labelled web-only in the builder so it is clear before you ship.

Stored credentials are kept for use in the headers you configure and are never written to logs — request and response logging redacts authorization headers and any known secret values automatically.

Yes — function calling, AI function calling and tool calling all describe the same idea: letting a model invoke a function you have defined and use the result in its answer. Robofy's Function Tool Call is that capability applied to your own HTTPS APIs, with the builder, the mandatory pre-save test, the encrypted credential store and the security guardrails wrapped around it. If you have used function calling through a model provider's API directly, the concept is identical; the difference is that here you don't write or host the calling code.