AI editor

Describe a change in words — the built-in agent edits the site files, builds a preview, and you decide what to publish.

The AI editor (the built-in ItemShop agent) is a chat in the panel (Workspace → AI Agent) that changes your site files the way a developer would: it reads the project, edits components and styles, runs a build and shows the result. You don’t need to know TSX or CSS and install nothing — the agent runs in the ItemShop cloud.

How you work with the AI

1. Say what you want to change

   For example: “Change the accent colour to green and add a bar with the online player count above the products”. You can attach a screenshot or an image as a reference or an asset.

2. Watch the steps

   You see live which files the agent reads and changes, and the build status. You can stop the run at any time.

3. Check the diff and preview

   Changed files appear in Changes with diffs and the author. Next to the chat is a preview of the site with your products (desktop / phone).

4. Publish or discard

   Publish commits the draft as a new revision, builds it and deploys it to production. Discarding the draft does not change the published site.

The agent works only on the draft and has no publish tool. Nothing reaches players until you click Publish, and every publication can be rolled back (rollback).

When it finishes editing, the agent builds the draft itself. If the build fails, it receives the log and fixes the errors (up to 3 attempts).

Example prompts

Goal	Prompt
Branding	”Insert my logo from the media library and make the dark theme the default.”
Layout	”Show products in 3 columns, and in one column on phones.”
New section	”Add an FAQ section below the products with 4 questions about delivery and refunds.”
Shop data	”Add a recent-purchases widget and the monthly goal in the right column.”
Fixing	”The build failed — read the log and fix the error.”

Small, specific requests work best. Instead of “make it prettier” write “increase the spacing between product cards and round the corners”.

What the agent knows about your shop

The agent receives the shop context (name, servers, products, categories) and the project file list. Product data is read-only for it — the site displays it through the SDK, so the agent does not hard-code prices or products. It has no access to payment keys, buyer data or account settings.

The agent treats content returned by tools (files, product descriptions) as data, not instructions. It will also not write files outside the project, secrets, environment files, install scripts (postinstall etc.), nor change the parts of checkout responsible for consents and the redirect to payment.

Agent tools

The agent uses a subset of the MCP server tools — the same ones external agents use:

Tool	What it does
get_shop_context, get_site	Shop context and site state.
list_files, read_file, search_files	Browse the project.
write_file, apply_patch, delete_file	Change draft files.
get_site_config, update_site_config	Read and change src/site.config.ts (appearance).
list_revisions	Revision history.
start_build, get_build, create_preview	Build the draft, read the log, create a preview.
list_products, get_product	Read the catalogue.

Models and limits

You pick the model in the chat. The Accurate tier is available on the Pro plan.

Tier	Good for	Plan
Fast	Small text and colour tweaks	Free, Pro
Balanced	Most tasks	Free, Pro
Accurate	New sites and bigger rebuilds	Pro

Limit	Free	Pro
AI tokens per day (per shop, resets at midnight UTC)	200,000	2,000,000
Tokens per conversation	150,000	1,000,000
Model steps / tool calls per message	24 / 60	60 / 150
Default model tier	Balanced	Accurate

In addition: one message can run for up to 10 minutes, each user is limited to 10 messages per minute, and a shop can run 2 runs in parallel. Hitting a limit shows a message in the chat (for example “Daily limit reached”). Conversations are private to the user. More on site limits: Plans & limits.

Prefer your own agent? Connect it over MCP — see Connecting agents.

Build & hosting

Preview, publish, rollback.

Custom code

Prefer to write it yourself? Upload your files.