Commerce that runs itself

AI agents can reason and write. They can’t transact. They’re trapped in a world built for humans with credit cards and web browsers — every order, refund, and inventory check goes through an API designed for a person clicking a button. StateSet is the infrastructure layer that fixes that. We build the engine agents run on and the rails they transact over, so an agent can buy, sell, and operate on its own.

I'm building an agent

Embed a full commerce engine in your runtime. Orders, inventory, payments, returns — local, typed, and agent-callable.

I'm running operations

Deploy agents that resolve customer work end to end. Priced per outcome, not per seat.

What we built

Layer	What it is	Why it’s unusual
iCommerce	An embedded commerce engine — “SQLite for commerce”	You don’t call it over a network. It runs inside your agent’s process
ResponseCX	AI agents that resolve customer work	They act, not just answer — and you pay per resolved outcome
NSR	Neuro-symbolic decision engine	Returns a machine-checkable proof, not a confidence score
ACP / ICP	Agent commerce protocols	”HTTP for money” — signed mandates, offline-verifiable receipts
Computer Use	Agents operating software UIs	Automates legacy tools that have no API at all

Plus the operational spine commerce actually needs: EDI for trading partners, Voice for phone, Mail for email, Sync across 180+ platforms, and a Temporal engine that makes every automated action durable and replayable.

Why it’s different

We’re an engine you embed, not an API you call. Stripe and Shopify hand you endpoints over the network. iCommerce compiles into your runtime — zero network latency, offline-capable, type-safe, with the full order and inventory state machine local to the agent. Our agents act, they don’t suggest. Copilots draft a reply for a human to send. StateSet resolves the ticket and closes it. That’s what makes per-outcome pricing possible instead of per-seat. Adaptive, not brittle. RPA breaks when a button moves. Our agents use vision and reasoning to adapt the way a person would. Consequential decisions come with proofs. Most AI platforms skip this. When an agent issues a refund or grants access, NSR returns approved, denied, or refused with the cited rules behind it. An approval only executes when its proof is grounded; if the policy engine is unreachable, the action escalates to a human rather than proceeding. A confidence number is not an authorization.

The full case — differentiation, value, and the hard questions

How we compare to Stripe, RPA, and copilots, what it costs, and honest answers on trust, lock-in, and where AI still gets things wrong.

Start building

The engine is open source, dual-licensed MIT OR Apache-2.0. Inspect it, fork it, run it anywhere.

npm install -g @stateset/cli
stateset --version

Or embed the engine directly:

npm install @stateset/embedded

Give an agent the whole surface — 719 MCP tools across 63 domains:

curl -s https://docs.stateset.com/stateset-icommerce-skill.md \
  > ~/.claude/skills/stateset-icommerce/SKILL.md

iCommerce Quickstart

Run the embedded engine.

MCP Servers

Connect any service to Claude.

API Reference

Every endpoint.

Explore by product

iCommerce

The embedded commerce engine — 11 language bindings.

ResponseCX

Outcome-priced AI agents for CX and commerce ops.

Temporal Engine

Durable, replayable automation in Rust.

NSR

Verified decisions with cited proofs.

Voice

AI voice agents over Twilio and OpenAI Realtime.

EDI

X12 and EDIFACT for trading partners.

Sandbox

Isolated execution environments for agents.

Sequencer

Verifiable event sequencing.

Set L2

On-chain anchoring of commitments and proofs.

Getting started for AI agents

​Commerce that runs itself