AI Voice Agent Backend

After this module you can give an AI phone receptionist (Retell AI or Vapi) real abilities: check availability, book, cancel and recognise callers, plus a summary of every call.

What it does

Voice platforms handle the talking; they call your webhook whenever the agent needs to do something. This workflow is that backend, for a demo clinic, Bright Smile Dental:

• POST /webhook/voice/tools: one endpoint for every tool: check_availability, book_appointment, cancel_appointment, lookup_customer. It accepts Retell ({name, args, call}), Retell's args-only mode (?tool=), and Vapi (message.toolCallList), and replies in each platform's format in ~25 ms. Every reply has a message the agent can read out loud.
• POST /webhook/voice/call-ended: receives the end-of-call webhook, summarizes the transcript with AI (intent, follow-up needed), and branches to an owner alert.

Bookings live in the workflow's static data, so the whole tool flow works with zero credentials. The Business tier swaps in Google Calendar.

Flow

Voice Tools API → Business Settings → Normalize Tool Calls → Route Tool ─┬→ check_availability ─┐
                                                                         ├→ book_appointment ───┤
                                                                         ├→ cancel_appointment ─┼→ Format for Platform → Respond
                                                                         ├→ lookup_customer ────┤
                                                                         └→ Unknown Tool ───────┘
Call Ended → Normalize Call → Summarize Call (OpenAI) → Call Record → Needs Follow-up? → Notify Owner / Log Call

Verification status (n8n 2.43.6)

• Tools endpoint: 16/16 checks pass without credentials (Retell, Vapi and args-only payloads; booking, double-booking, cancel, lookup, weekend/past/missing-field handling, unknown tool). Median response 24 ms.
• Call-ended webhook: acknowledges immediately; a real transcript reaches the AI node and stops at "Error in sub-node OpenAI Chat Model" until you add an OpenAI credential (WITH_CREDS=1 ./test/test.sh).
• Not tested against a live Retell/Vapi phone call. Payload formats follow their docs (Retell custom functions, Vapi custom tools). Log one real request before going live (lesson, Step 6).

Files

File	Tier	Purpose
lesson.md	Free	Step-by-step build + connecting Retell/Vapi
workflow/starter.json	Free	Endpoint, normalizer and availability tool
workflow/learner.json	Learner	Finished backend (tools + post-call summary)
workflow/code/	Learner	Readable copies of every Code node
exercises.md	Learner	Practice tasks
solutions.md	Learner	Exercise solutions
notes.md	Private	Learning log
sell/	Private	Offer, video script, blog post

The JSON workflow is the source of truth; workflow/code/*.js are copies for reading and diffing.