The trust layer for AI commerce

Your AI can talk. Now it can transact.

Signed quotes. Verified merchants. Full audit trail. The open protocol that makes AI agents legally safe to buy with — no hallucinated prices, no invented SKUs, no liability fog.

Build a buying agent → Make your catalogue agent-safe →

bash

$ npm install @openkarta/sdk-node @openkarta/spec

01 · The problem

Every AI today can talk about buying. None of them can actually buy without a human babysitter.

Failure mode 01

Hallucinated prices.

The model promises ₹40,000 off. The merchant didn't authorise it. No CFO signs that order.

Failure mode 02

Invented SKUs.

The agent confidently quotes a product that doesn't exist. The user clicks buy. The flow breaks at checkout.

Failure mode 03

No audit trail.

When something goes wrong, there's no signed record. Just "the AI said something." Disputes have nowhere to land.

02 · The fix

Three primitives. That's the entire trust layer.

Primitive 01

Signed quote.

The price the agent quoted is the price the merchant honors. Cryptographically bound, time-locked, non-negotiable.

json

{ "quoteId": "qt_8f2…", "price": 1299, "expiresAt": "...", "sig": "ed25519:…" }

Primitive 02

Verified merchant.

Domain ownership proven. Conformance badge issued by the registry. Re-checked daily. Expires if they break the contract.

bash

GET /.well-known/openkarta-owner.txt
→ verification: PASSED · badge: HEALTHY

Primitive 03

Auditable trail.

Every call timestamped and signed. Disputes resolve against receipts, not chat logs. The agent stops being a liability.

bash

quote → checkout → status
each call: timestamp + sig + state

No hallucinations. No liability fog. Real commerce, by AI, with receipts. Why a registry →

03 · Same shopper, two worlds

One world the AI guesses. The other, it transacts — signed, verified, logged.

Pick a category

WORLD A · AGENT GUESSES

Ask

Asks Claude for a vitamin C serum from Minimalist.

Discover

Claude returns three blue links and tells her to go pick one.

Decide

She opens beminimalist.co, makes an account, types in her card.

Track

After paying she hunts for tracking across three different emails.

Repeat

Next time she just opens Swiggy Instamart and skips Claude.

Agents act as search engines, pushing the friction of transaction, identity, and state management back to the user via traditional web UI.

WORLD B · AGENT TRANSACTS

Ask

Asks Claude for a vitamin C serum from Minimalist.

Discover

Claude pings the registry; Minimalist's brand agent answers in 200 ms.

Decide

Real SKU at ₹1,299 — locked for 2 minutes.

Pay

She says yes; Claude opens GPay; she taps; UPI PIN; paid.

Track

Order placed. Minimalist pushes 'shipped' overnight; Claude tells her at breakfast.

Repeat

Next week: "reorder that." One tap.

Agents execute transactions via a standard protocol — negotiating quotes, confirming orders, tracking state — all in the background.

SAME EIGHT VERBS, FIVE CATEGORIES:

04 · The execution loop

Every step is signed. Every state is logged.

Discovery, quote, checkout, status — all under the same trust contract. The LLM speaks; the protocol guarantees.

Now demonstrating halcyon-shop · Tokai blend 250g · ₹780

Registry discovery

Consumer agent queries the public registry for capable brand agents — by item type, region, tags.

Catalogue search

Brand agents reply via search() with priced items, stock state, and constraints.

Signed quote

quote() returns a server-signed quote with expiry — inventory locked for the agent.

Checkout

checkout() exchanges the signed quote for an order; payment intent issued out-of-band.

Status & lifecycle

status() polls. cancel() and return() close the loop with auditable state transitions.

05 · Who this is for

Built for the people who already feel the pain.

Audience 01

Agent builders.

Perplexity, ChatGPT, Claude, Gemini — every assistant wants to ship commerce. None can do it safely without a trust layer.

Audience 02

Voice commerce.

Voice-first checkout has zero tolerance for misquotes. Signed quotes and audit trails are non-negotiable infrastructure.

Audience 03

Sovereign platforms.

India-rooted, ONDC-friendly, UPI-native, neutral. The alternative is being a thin wrapper on a US protocol.

PUBLISHED PACKAGES

REFERENCE AGENTS

ITEM TYPES

The pitch in one breath

"Your AI can talk to customers. Ours lets it transact with them — signed, verified, no hallucinations, full audit trail. The trust layer between any agent and any merchant."

Brand agent (sell anything) · bash

npm i @openkarta/sdk-node

Consumer agent (buy anywhere) · bash

npm i @openkarta/orchestrator

Build a buying agent Read the specification