Run SchemaSmith from Claude Code
You’ve run the loop by hand: install, connect, extract, preview, deploy. If you already work in Claude Code, there’s a second way to drive the same tools. Same switches, same settings, same exit codes. You talk, it runs the command, and it tells you what came back.
Install the plugin
Section titled “Install the plugin”From 2026.12.0, the SchemaSmith repository carries a Claude Code plugin. Two commands in Claude Code:
/plugin marketplace add Schema-Smith/SchemaSmith/plugin install schemasmith@schemasmithIt works in Claude Code’s terminal, the IDE extensions and the desktop app’s Code tab. Its version moves with the tools’, so the plugin and the CLI it drives always agree on what a switch means.
Eight skills, one loop
Section titled “Eight skills, one loop”Each skill is one step of the loop you already know:
| Skill | What it does |
|---|---|
/schemasmith:install | Installs the tools for your operating system and checks each answers --version |
/schemasmith:connect | Writes a settings file for a server and proves it with --TestConnection |
/schemasmith:extract | Extracts a database into a new schema package and validates it |
/schemasmith:preview | Validates a package and previews a deployment with --WhatIf, in plain words |
/schemasmith:deploy | Runs a deployment after a preview, once you confirm the server and database |
/schemasmith:drift | Checks whether a database still matches its package |
/schemasmith:explain-run | Reads the last run’s exit code, summary and logs, and says what happened |
/schemasmith:ci | Writes a GitHub Actions workflow using the SchemaSmith Deploy action |
Call one by name, or just ask: “extract my Shop database into a package”, “what would a deploy change?”, “why did that fail?”
What it does with your server
Section titled “What it does with your server”The plugin runs the tools through Claude Code’s own shell, so you see every command before it runs. The order is fixed, and it’s the same order Module 3 taught you.
Connect proves the connection first. It asks for the platform, the server, the port and the user, writes SchemaQuench.settings.json in your working folder, and runs SchemaQuench --TestConnection. That connects and checks the server’s version against the package’s minimum. It changes nothing. If a settings file is already there, it shows you what would change and asks before replacing it.
Preview validates before it predicts. --Validate runs first. If the package has errors, you get them with their SS- codes and the preview stops there, because a preview of a broken package tells you nothing. Then --WhatIf, and a plain-words read of the summary: what each database would gain, change and lose, which scripts would run, and any change that would remove data, with the setting that governs it.
Deploy asks every time. No preview of this package against this target in the conversation? It runs one first. Then it shows the exact command, names the server and the databases it’ll change, and waits for your yes. That yes covers that target only. A different server gets its own preview and its own yes. And it never changes a safety setting to get a deployment through: if the destructive-change policy would refuse something, it tells you which setting governs it and leaves the call with you.
Preview, name the target, wait for a yes. That’s the same discipline you’d want from anyone you handed the keys to.
What it doesn’t do
Section titled “What it doesn’t do”It doesn’t edit your package. Tables, scripts and templates are yours to write, with the package reference and the JSON schemas your editor reads. After any edit, it runs SchemaQuench --Validate for you. The package stays yours, and validation stays the gate.
It never asks for a password. Passwords come from environment variables you set in your shell before you start Claude Code, SmithySettings_Target__Password for SchemaQuench and SmithySettings_Source__Password for extraction. They never go in a settings file, and never into the conversation. On SQL Server with Windows authentication there’s no password to set.
One thing it can’t avoid: extraction, preview and deployment install SchemaSmith’s own helper objects on a writable target, exactly as running the tools by hand does. Extraction and preview change nothing of yours.
Try it on the sandbox
Section titled “Try it on the sandbox”There’s no lab for this lesson, because a lab here is a conversation and every conversation reads a little differently. But you can run the whole loop against the same chinook database Module 4 brought under management. Load it first (Module 4, Step 1), with the sandbox up.
Set the passwords in the shell you’ll start Claude Code from. The plugin reads them from there and never asks:
export SmithySettings_Source__Password='Learn!Passw0rd'export SmithySettings_Target__Password='Learn!Passw0rd'Start Claude Code in an empty folder, install the plugin, and ask, one at a time:
/schemasmith:install. If the tools are already on your PATH, it checks each answers--versionand moves on.- “Connect to my SQL Server at localhost, port 11433, user sa.” It writes
SchemaQuench.settings.jsonand runs--TestConnection. The sandbox’s certificate is self-signed, so expect it to point you atTrustServerCertificate. (PostgreSQL is port 15432 with userpostgres; MySQL 13306 and MariaDB 13307 with userroot.) - “Extract the chinook database into a package.” That’s SchemaTongs, then
--Validateon what it wrote. - “What would a deploy of this package to chinook change?” That’s
--Validate, then--WhatIf, read back in plain words. You just extracted it from that database, so the answer should be: nothing. - Change something in the package yourself, a column’s length say, and ask again. Now the preview has one change. Ask it to deploy, and watch it name the server and the database and wait for your yes.
The words will differ from run to run. The commands won’t: they’re the ones the earlier lessons had you type.
The full page, with the same tables, is the Claude Code guide.
Check yourself: You previewed a deploy to your staging server and said yes. Now you ask the plugin to deploy the same package to production. What happens, and why?
It previews production first, then shows the command, names the production server and databases, and waits for a new yes. A yes covers the target it was given for and nothing else. A preview of staging says nothing about what production holds, so it can’t stand in for production’s preview.
The plugin doesn’t add anything to the tools. It’s the same forge with a second pair of hands at the bellows, and those hands show you every strike before it lands. You still own the package, you still read the preview, and nothing touches a server you didn’t name.
Questions, or a skill that told you something the tools didn’t back up? forgebarrett@schemasmith.com. I read every one.
Until then, may every command you’re shown be the one that runs, and every yes you give name its target.
— Forge