Skip to main content
Commerce Engine is designed for agent-assisted development. The recommended workflow combines three sources of truth:

Agent skills

Decision trees, architecture rules, implementation patterns, and known pitfalls for complete commerce features.

LLM-first reference

Exact, current SDK methods, operation examples, request types, response types, and schemas retrieved only when needed.

Production starters

Working storefront implementations across supported frameworks, including rendering, performance, SEO, session, and deployment patterns.

​
Install the skills package

Run this command in the storefront repository:
The package installs a Commerce Engine router and task-specific skills for setup, authentication, catalog, cart and checkout, orders, webhooks, first-party SSR frameworks, and custom SSR bindings.
Start broad Commerce Engine tasks through the ce router when your agent supports skills or slash commands. The router identifies the framework and loads the relevant specialist guidance.
For Claude Code, you can also install the plugin marketplace:

​
The agent workflow

A reliable agent should follow this sequence rather than generating from memory:
1

Inspect the project

Read package.json, framework configuration, existing routes, styling conventions, package manager, deployment configuration, and environment-variable patterns.
2

Load the relevant skills

Load setup plus every feature skill needed by the request. A complete storefront normally spans framework setup, catalog, Hosted Checkout, authentication, and orders.
3

Retrieve exact contracts on demand

Before writing an SDK call, fetch the current operation page or schema from the LLM-first reference. Do not guess method signatures or duplicate API interfaces.
4

Use a starter as implementation evidence

Compare framework boundaries, bootstrap, session ownership, rendering, SEO, loading states, and deployment configuration with the closest production starter.
5

Implement and validate

Run the repository’s typecheck, lint/check, and production build commands. Exercise the relevant customer journeys and report anything that could not be verified.

​
Context-efficient SDK lookup

The LLM reference is available at llm-docs.commercengine.io. Every page supports raw Markdown through content negotiation or a .md suffix.
Equivalent .md URLs work in agents that cannot send custom headers:
This progressive-disclosure model keeps context small: load the architectural skill first, then retrieve only the operation and schemas required for the feature being implemented.

​
Current SDK mental model

All supported frameworks install the unified package:
The core rule is:
  • Public reads use the public accessor. They are publishable-key-backed and do not create a user session.
  • Live user flows use a session accessor. Authentication, cart, wishlist, account, orders, and payments belong to the anonymous or logged-in session.
  • Hosted Checkout should use authMode: "provided" with two-way token synchronization whenever the application also uses @commercengine/storefront.
Framework wrappers expose the same distinction with context-appropriate accessors:

​
Type lookup rule

Import Commerce Engine schema and operation types directly from @commercengine/storefront. Do not recreate API interfaces in application code.
Operation types follow a predictable naming scheme:
  • {OperationName}Body
  • {OperationName}Content
  • {OperationName}Query
  • {OperationName}PathParams
Use the LLM-first operation page to confirm the exact exported names before importing them.

​
Build a complete storefront

Production storefront workflow

A complete agent execution contract, route matrix, validation gates, and copy-paste prompt.

Production starters

Framework-by-framework reference implementations and the patterns agents should study before generating code.

Prompt library

Prompts for new storefronts, existing-project integrations, feature work, debugging, and production review.

LLM API/SDK reference

Browse exact methods, operations, schemas, webhook payloads, parameters, and responses.
Skills and generated code do not replace validation. An agent should not claim a storefront is production-ready until the project builds successfully and the implemented customer journeys have been exercised.