claude-pizza

A Claude Code skill + CLI that orders a pizza for you over a real phone call, end-to-end via voice.

Trigger from Claude Code with a natural-language command. The skill:

1. Calls you via voice agent and asks what pizza you want (in Italian).
2. Searches the web for a pizzeria near your address that's open now.
3. Calls the pizzeria via voice agent and places the order on your behalf.
4. Reports back to Claude Code (transcripts + outcome).

No daemons, no scheduling. One command, one order.

---

User flow

[Paolo in terminal]
  $ claude
  > Ordina una pizza

[Claude Code]
  invokes skill `pizzaiolo`
  → spawns CLI: pizza order

[CLI step 1 — call Paolo]
  Vapi outbound call → PAOLO_PHONE
  Agent (IT): "Ciao Paolo, che pizza vuoi stasera?"
  Paolo (voice): "Una diavola e una margherita, consegna a casa"
  Agent confirms, hangs up
  → structured order JSON returned via Vapi

[CLI step 2 — find pizzeria]
  Google Places API: type=restaurant, keyword=pizzeria,
    location=HOME_ADDRESS, radius=1500m, opennow=true
  Rank by rating × review_count, filter has_phone_number
  Pick top 1 (with 2 fallbacks)

[CLI step 3 — call pizzeria]
  Vapi outbound call → pizzeria number
  Agent (IT): scripted order, asks for total + ETA,
    gives delivery address, confirms
  → transcript + outcome returned

[CLI step 4 — report]
  Print summary to terminal:
    ✓ Ordered from Pizzeria Spontini (via Padova 12)
    ✓ 1× Diavola, 1× Margherita
    ✓ Total: €18, ETA: 35 min
    ✓ Confirmed by: Marco (pizzeria)
  Save full transcripts to ~/.pizzaiolo/orders/<timestamp>.json

If the pizzeria call fails (no answer, refuses, doesn't deliver to address), the CLI falls back to the next candidate. Max 3 attempts, then the order is reported as failed.

---

Quick start

# 1. Install dependencies
npm install

# 2. Copy env template and fill in your keys
cp .env.example .env
$EDITOR .env

# 3. Build
npm run build

# 4. Place an order
npm start order

# Or run from source without building
npm run dev order

For development, use --dry-run to test the full flow without real phone calls:

npm run dev order -- --dry-run

---

CLI options

Flag	Description
--dry-run	Simulate the full ordering flow without making real Vapi calls. Google Places is still called for real.
--skip-ask	Skip calling Paolo and reuse the items from the last saved order.
--candidate <n>	Pick the Nth pizzeria from search results (1-based) instead of the top match.
-v, --verbose	Verbose debug output for troubleshooting.

Examples:

pizza order                       # full live run
pizza order --dry-run             # safe test, no calls
pizza order --candidate 2         # use 2nd-ranked pizzeria
pizza order --skip-ask            # reuse last order
pizza order --verbose             # debug mode

---

Architecture

Pure local CLI. No server, no cloud webhooks — the CLI polls Vapi for call completion.

┌─────────────────────────────────────────────┐
│  Claude Code (terminal)                     │
│    user: "Ordina una pizza"                 │
└──────────────────┬──────────────────────────┘
                   │ invokes skill
                   ▼
┌─────────────────────────────────────────────┐
│  pizza CLI (Node.js + TypeScript)           │
│  ┌──────────────┐  ┌──────────────┐         │
│  │ OrderTaker   │  │ Locator      │         │
│  │ (calls you)  │  │ (Places API) │         │
│  └──────┬───────┘  └──────┬───────┘         │
│         │                 │                 │
│         └────────┬────────┘                 │
│                  ▼                          │
│          ┌──────────────┐                   │
│          │ Placer       │                   │
│          │ (calls shop) │                   │
│          └──────┬───────┘                   │
│                 ▼                           │
│          ┌──────────────┐                   │
│          │ Reporter     │                   │
│          └──────────────┘                   │
└─────────────────────────────────────────────┘
         │              │
         ▼              ▼
      Vapi API    Google Places
    (voice agent)     API

---

Components

OrderTaker — calls Paolo (src/ordertaker.ts)

• Vapi outbound call to PAOLO_PHONE.
• Voice agent prompt (Italian): friendly, brief, asks pizza type(s), size, delivery vs pickup, any notes.
• Returns structured output:

  {
    "items": [{ "name": "Diavola", "size": "media", "notes": "" }],
    "delivery": true,
    "user_confirmed": true
  }

• If user_confirmed=false or Paolo cancels → exits cleanly.

Locator — finds pizzeria (src/locator.ts)

• Google Places API nearbySearch.
• Query params: type=restaurant, keyword=pizzeria, location=HOME_ADDRESS, radius=1500, opennow=true.
• Filters: business_status=OPERATIONAL, rating >= 4.0, has phone number.
• Ranks by rating * log(user_ratings_total).
• Returns top 3 candidates with {name, phone, address, rating}.

Placer — calls pizzeria (src/placer.ts)

• Vapi outbound call to candidate's phone.
• Voice agent prompt (Italian): identifies as "ordine per Paolo", reads items, confirms delivery address, asks total + ETA, confirms.
• Returns structured output:

  {
    "accepted": true,
    "total_eur": 18.0,
    "eta_minutes": 35,
    "operator_name": "Marco",
    "rejection_reason": null
  }

• If accepted=false → tries next candidate (up to 3 total).

Reporter (src/reporter.ts)

• Pretty-prints result to stdout.
• Saves full payload — including both call transcripts — to ~/.pizzaiolo/orders/<ISO>.json for debugging.

---

Tech stack

• Runtime: Node.js 20, TypeScript
• CLI framework: commander
• Voice: Vapi — outbound calls + structured extraction, Italian support.
• Pizzeria search: Google Places API (v1).
• Config: .env at project root.
• Persistence: flat JSON files under ~/.pizzaiolo/. No DB.

---

Configuration

Copy .env.example to .env and fill in:

VAPI_API_KEY=                    # from https://vapi.ai/dashboard
VAPI_PHONE_NUMBER_ID=            # Vapi-provisioned IT number for outbound
VAPI_ORDERTAKER_ASSISTANT_ID=    # pre-built in Vapi dashboard
VAPI_PLACER_ASSISTANT_ID=        # pre-built in Vapi dashboard

PAOLO_PHONE=+39...               # E.164 format
HOME_ADDRESS=Via ..., Milano

GOOGLE_PLACES_API_KEY=

All variables are required — the CLI fails fast on startup if any are missing.

---

Claude Code skill

The repo ships a Claude Code skill at .claude/skills/pizzaiolo/SKILL.md. It triggers on phrases like:

• ordina una pizza, ordiniamo la pizza, voglio una pizza
• ho fame, prendi una pizza, pizza time
• order a pizza, get me a pizza

Dictation → text → skill match → runs pizza order.

---

Cost estimate per order

Item	Cost
Vapi call to Paolo (~1 min)	~€0.15
Vapi call to pizzeria (~2 min, possibly 2 retries)	~€0.30–0.90
Google Places (search + details)	~€0.03
Total	~€0.50–1.10 per order

Plus, you know, the pizza.

---

Risks & open questions

1. Italian pizzaiolo on the phone is the hard eval. Expect to iterate the Placer prompt heavily. First few orders: use --dry-run and call yourself pretending to be the pizzeria to test the agent.
2. Phone number quality from Google Places is mixed — some entries have outdated numbers. Mitigation: cross-check with the pizzeria's website if listed.
3. Payment: most Italian pizzerie are cash/card on delivery. The agent doesn't handle payment, just places the order.
4. Legal: outbound automated calls in IT — calling a business to place a legitimate order is fine. Do not point this at consumers.
5. Voice triggering: Claude Code's skill matching is text-based. Dictation → text → skill match works as long as the skill description includes the Italian trigger phrases.

---

Out of scope (v1)

• Scheduling / cron pings
• Multiple recipients
• Tracking the delivery
• Anything other than Milan
• English-language pizzerias
• Payment handling
• Reordering UI