Driving a migration with your AI assistant

Everything you do on a migration in the dashboard — looking at source data, adjusting mappings, starting and watching runs, and checking what was loaded — can also be handed to an AI assistant acting as a migration engineer for you: it investigates your data, adapts the mappings to your requirements, proposes runs, watches them, and tells you when something looks wrong.

There are two ways to connect one. You stay in control of anything that changes your migration either way.

Connect Claude or ChatGPT (no install)

Graftport is available as a connector — the built-in way AI apps plug into outside services. Nothing to install, no keys to copy:

1. In Claude: Settings → Connectors → Add custom connector, and paste the Graftport connector URL: https://api.graftport.com/mcp. (In ChatGPT and other MCP-capable apps, add the same URL wherever they take custom connectors.)
2. Click Connect. Your browser opens the normal Graftport sign-in, then a single approval screen showing exactly what the assistant will be able to do.
3. Click Allow — you’re connected. Ask the assistant to “list my Graftport migrations” to see it work.

The assistant acts as your Graftport account: it sees only your migrations, and its built-in instructions teach it the whole mapping workflow — no setup on your side.

Or use the graftport CLI (for coding agents)

Developers who work in a terminal-based coding agent (Claude Code, Cursor, Windsurf, …) can use the graftport command-line tool instead — same abilities, driven over the shell. See Getting started below.

What the agent does on its own

The agent can do everything that doesn’t cost money without stopping to ask:

• Browse your migrations, resources, and mappings.
• Understand a resource that has more than one source — your store’s own data plus any CSVs you’ve added — and work on the right mapping for each source, since every source has its own.
• Sample real source records (both the cleaned-up shape and the raw data from your old store), for one source at a time when a resource has several.
• Edit a mapping, validate it against real records until every sampled record passes, and publish the new mapping version so it’s ready for the next run.
• Run a dry-run — the agent can ask Graftport to compute exactly what would be sent to Shopify, without actually sending it. That way it can read the result and tell you whether the migration is ready before you pay for a real run.
• Read run progress, group load errors so the biggest problem is obvious, and inspect what was actually loaded.

What always needs your approval

Anything that costs money on the platform or could destroy in-flight work is proposed, not performed. The agent composes the exact action and shows it to you; it runs only after you say yes:

• Starting a real run (anything that actually pushes to Shopify). You see a preview of exactly what it will do and the estimated cost.
• Cancelling a run.
• Retrying a failed record.

The agent is instructed never to skip this confirmation — and for an assistant connected through the connector the guarantee is even stronger: those actions simply don’t exist on its side. It prepares the action and points you at the dashboard; only you can pull the trigger.

A “passing” run is not always a correct run

A run can finish green and still have migrated something subtly wrong — empty product descriptions, collapsed variants, prices that don’t add up. The agent is taught to inspect the loaded result against a per-resource checklist and report anything suspicious instead of declaring success on a green check alone. Treat its summary as a second pair of eyes, not a guarantee — review the result in Shopify before going live.

Getting started with the CLI

Install the tool and sign in. auth login opens your browser, signs you in the usual way, and asks you to authorize the CLI on a single “Authorize Graftport CLI” screen. Nothing to copy and paste.

graftport auth login
graftport skill install

skill install drops an up-to-date instruction guide into your AI agent so it knows how to use every command and, importantly, when to stop and ask you. From there, tell your agent what you want — for example, “prepare the product migration so prices come from the tier price when the base price is missing” — and it will work the loop, pausing for your approval before any change.

If you are on a headless server with no browser, pass --no-browser to print the sign-in URL so you can open it on another device, or --access-token <jwt> if you already have a token from your dashboard session.

If you’d rather drive it yourself, run graftport --help; every command also has --help, and the dashboard does the same things through buttons.