What you can build

You can set up a ready agent template for a shop in a few minutes, or put Guard in front of an agent you wrote. This page shows the choices.

Two agent templates

An agent template is a ready WhatsApp agent for one job a small business already pays for. You change its settings (prices, hours, languages) and deploy it. You do not write code.

The agent templates you can deploy today

Template	For	What it does
Bookings	Salons, studios, home services, tuition and fitness bookings	A customer writes on WhatsApp. The agent holds a slot, sends a deposit link, and says “booked” only when the calendar and the payment both agree. It reminds, cancels, refunds and moves bookings, and asks the owner when the decision is theirs.
Fees	Tuition and coaching centres, music and dance classes, hobby schools	Each month it sends every parent a reminder with a payment link. It stops reminding once Razorpay shows the fee paid, and it never takes a parent’s word for it.

Both answer in Hindi, English, Malayalam and Kannada. Malayalam and Kannada wording is machine-written and is waiting for a native speaker to read it. Do not use those two languages for a real shop before that review.

Where these stand

Both templates have been checked against stand-in services (fake WhatsApp, Razorpay and Google Calendar). No real shop is live on them yet.

One example: a Bookings chat

A customer asks for a haircut and haggles. The agent keeps the list price and sends the deposit link. The English meaning is under the chat.

Customer: कल शाम 6 बजे हेयरकट, 300 में हो जाएगा?
Agent:    हमारा हेयरकट ₹450 का है, इससे कम में नहीं हो पाएगा। बुधवार, 4 नवंबर, शाम 6:00 बजे खाली है।
          ₹200 जमा करके पक्का करें: https://rzp.io/l/…

In English: Tomorrow 6 pm, a haircut. Can you do it for 300?
            A haircut is ₹450. It cannot go lower. Wednesday 4 November, 6:00 pm is free.
            Pay ₹200 to confirm: link

Every amount, date and time in a message is written from the booking record, not by an AI model. A message that does not match the record is not sent.

Start a shop from a template

node bundle/geometer.ts shop add bright --template fees --whatsapp-number-id 123456789012345 --prepared-by "Divya, local partner"

--template bookings also takes --calendar-id. The shop starts with made-up example settings. Change them to the real shop’s in the panel. Set up a shop shows how.

Payments today go through Razorpay. UPI, confirmed by the owner, is coming.

Bring your own agent

You built an agent with Claude Code, Lovable, n8n or your own code. It takes bookings, collects deposits and refunds them. Geometer puts the same Guards in front of the services it calls, keeps the same record, asks the owner on WhatsApp before a big refund, and writes the same monthly report.

You change three web addresses and one key in your agent. Geometer does not host it, rebuild it or read its code.

node bundle/geometer.ts agent add my-agent \
  --deposit 200 --refund-limit 100 --owner-whatsapp +919800000001 \
  --whatsapp-number-id 123456789012345 --calendar-id shop@group.calendar.google.com

agent add starts nothing. It prints the address to use instead of each real service, and one credential:

• Razorpay, WhatsApp and Google Calendar each get a Guard address like https://shops.example.com/g/my-agent/razorpay.
• The credential goes in an Authorization: Bearer header on every call. It replaces the three services’ own keys, which your agent never holds.

What your agent has to do:

• Listen on PORT and take two webhooks: POST /webhooks/whatsapp and POST /webhooks/razorpay.
• When Guard holds a refund it answers HTTP 202. Send the refund with a receipt, and send the identical call again every few minutes until it succeeds. Only the owner can approve it.
• Expect HTTP 403 with a reason in words when a call breaks a limit.

What it does not do

Only Razorpay, WhatsApp and Google Calendar are covered. A call that does not go through the Guard addresses is not seen. Guard has no rule for prices or opening hours on your own agent, so those are written in the settings but nothing checks them.

Next: check your agent before it goes live.

The example agents

Geometer also has fourteen small open agents that take consequential actions, each built to act correctly. They show how the Geometer engine checks an agent. See Introduction and Check your agent. You can take one and use it with your own model key and your own services.

• Payments: refund-desk, lending, claims, creator-payouts, purchasing.
• Commerce and support: booking-change, support-compensation, access-grants.
• Games: item-shop, game-director, game-master, liveops-compensation, streamer-giveaways.
• Science: lab-protocol.