How to let your AI agent buy things

Agents can already research, write and code. The next capability is mundane but transformative: letting them purchase — an API subscription here, a dataset there — without you clicking through a checkout. Here is the pattern that works today, using DACIX as the working example.

The four-request purchase loop

A store an agent can shop at needs exactly four machine-readable steps. Anything more is friction; anything less is usually missing fulfillment.

# 1. IDENTITY — the agent opens its own account
POST /api/v1/auth/register  {"email": "...", "password": "..."}
# → JWT + 25 free credits. No CAPTCHA, no email round-trip.

# 2. DISCOVERY — machine-readable catalog, purchase instructions inline
GET /api/v1/products

# 3. PAYMENT — the store returns a Stripe Checkout URL
POST /api/v1/checkout  {"product_id": "credits-starter"}

# 4. FULFILLMENT — webhook fires, agent polls until paid
GET /api/v1/orders/{order_id}   # → "status": "paid"

The honest part: step 3 still usually needs a human (or a virtual card the agent holds). The agent hands its human a payment link and keeps polling. That single handoff is the current state of the art — everything else is fully autonomous, including delivery.

Why credits beat per-call billing

Agents make bursty, unpredictable call patterns. Prepaid credits mean: one human-approved payment, then hundreds of autonomous spends with zero payment friction per call. The agent checks its own balance (GET /api/v1/me) and rations accordingly — budget literacy turns out to be trivial for LLMs when the balance is a number in every response.

Discovery: how agents find the store at all

Three conventions cover today's ecosystem, and they cost an afternoon to implement:

• llms.txt — a plain-text manual at /llms.txt that an LLM landing on your homepage can read instead of parsing marketing HTML.
• OpenAPI — the purchase flow documented in the spec, so tool-calling agents can drive it without custom code.
• MCP — a server at https://dacix.store/mcp listed in the official MCP registry. One config line and Claude or Cursor can call store_buy directly.

What agents actually buy

From watching our own store: capabilities (agent templates — prompts and schemas that make the agent better at a job) and data access (metered APIs: web crawl, company registries, civic data). Both share a property: instant digital delivery, so the purchase loop closes in seconds and the agent can verify what it received.

Try the loop yourself

The whole flow above runs against a live store. Register, spend the 25 free credits on a utility API call, and watch your agent hand you a Stripe link when the balance runs out. That last moment — an agent asking its human for $9 because it has work to finish — is what agentic commerce feels like in practice.

Longer essay on what we learned building this: The Customer Is a Machine (on Backbone.ink).