Siteglide MCP (Alpha)

What is it?

Siteglide MCP is a catalogue of tools which your AI coding assistant can use to interact with your Siteglide project. As well as editing files locally, the assistant can validate Liquid, run GraphQL, evaluate templates against staging, and read logs — using the same Siteglide CLI setup you already use.

Works with Cursor, Claude Code, Windsurf, and other editors that support MCP.

Why use it?

Catch mistakes before they reach staging
The assistant automatically validates Liquid, GraphQL, and YAML against Siteglide rules as it goes along, catching and fixing problems in real time.

Test against real data
Your assistant can run Liquid and GraphQL against your site from the editor, reading and updating your data. Its ability to work with data makes it very powerful, but is a major reason why we recommend working on staging sites only while the MCP is in alpha.

Debug faster
Pull recent platform logs when something fails, instead of switching between the admin, terminal, and your IDE.

Safer by default
Credentials stay in Siteglide CLI on your machine. The assistant uses environment names (like staging) — it should never read or paste your config tokens.

Same platform as your site
MCP talks to the same environments you sync to with siteglide-cli sync. What the assistant validates is what you’ll see on staging.

What you need

• A site created in Siteglide (staging is strongly recommended while the MCP is in alpha)

• Siteglide CLI installed and an environment (e.g. staging) added

• A Siteglide project pulled locally (siteglide-cli pull)

• An MCP-capable AI agent on your machine (e.g. Cursor)

Setup

1. Pull your project

siteglide-cli pull <env>

This downloads your site files and writes MCP configuration for your editor.

2. Turn on the MCP server in your editor

In Cursor: open Settings → Tools & MCP, enable the siteglide server, then reload the window.

3. Confirm it’s working

Ask your assistant something like:

List my Siteglide environments and validate my homepage Liquid against staging.

If MCP is connected, it can list your sites and run validation while the network is offline. All other tools use the CLI to connect to the network.

How to use it

You don’t call MCP tools yourself — you describe what you want in chat, and the assistant uses the right tool when its description matches the task you’ve asked the assistant to complete. This helps you to be efficient with tokens.

What the assistant can do

Staging v.s. production safety

We think you’ll find using an AI assistant smooth and accurate when it gets its hands on the Siteglide MCP! Having said that, we want to ensure live sites are safe from accidental edits to code and data. While the MCP is in alpha, there are still some areas which can be improved, we’ll list both what we have done already to protect your site, and also be clear about the temporary limitations.

There may be other safeguards besides the ones we listed here. The purpose of listing these is to help you feel safe and secure that you understand how MCP works and you can judge your own appetite for risk.

Main limitations of safeguards:

• The alpha version of MCP is not very good yet at reading your terminal to determine if sync is already running, or if two different sync commands are running, possibly pointing at different environments. Some of the build in safeguards therefore have reduced effectiveness as code could be generated while sync is running and the assistant may not be aware (so be vigilant about this yourself for now!). We intend to fix this in our next release.

• Be careful to pull the site before the AI assistant generates code, just as you would for any project using Siteglide-CLI. We aim to improve tools to make collaboration easier in the next update.

Existing safeguards:

• Classification of environments Siteglide MCP has a specific tool it uses to check your environments e.g. staging and production. This helps keep your auth tokens safe and also means your assistant can tell which environment is a staging or production site even if the environments are confusingly named.

• Tool blocks on production

  • The liquid_exec tool which runs liquid on the fly, without writing to files, is blocked on production. Instead if the assistant needs to plan something like a data normalisation on production, it can write Liquid code which can be reviewed first, before being synced and run. (Again not yet recommended).

  • The assistant is allowed to use GraphQL to read data on production and can make you a report, but if its task requires it to update data on production without writing code, it has a strong rule that it must ask you first. It should be said that implementation of this rule then falls to the assistant to comply with. We strongly recommend that you say no while MCP is in alpha, but if you know what you’re doing, you can allow at your own risk.

• Data that is read from your site is passed through “untrusted data” tags to help the assistant ignore any malicious instructions which may be hidden in the data (this is a common problem across all AI tools and is partially implemented at the assistant level, rather than the MCP level alone).

• There is a policy that if the assistant detects a production sync is running, it will prompt you to decide how to proceed, before it generates any code. Note that this is a fallback, not something which can catch every time. As noted above, it can currently sometimes fail to spot a sync command running, but we will soon fix that limitation.

• The MCP does not have any tools telling it how to use Siteglide CLI sync, pull or deploy itself. This part of the workflow remains a human decision. However, the MCP does not currently explicitly block the assistant from using these commands if you tell it to.

Good practice tips

• We recommend turning on your assistant’s privacy mode or turning off using your data to train the AI assistant. Otherwise, there is a risk of sensitive data falling into the hands of others.

• If you would like a staging version of your site for testing, agencies can create a staging site copy in the copies tab.

• If you have more than one environment, e.g. a staging and a production site, say staging when you want live tests (safest default). This is a good safeguard in case the AI misinterprets where the changes should be made or where data should be read from.

• If siteglide-cli sync is running, confirm it’s pointed at staging, not production. If you absolutely must use your AI assistant on a production site, despite our recommendation not to while the MCP is in alpha, we recommend the following:

  • Turn sync off before asking the agent to generate code

  • Generate code

  • Review the code

  • Turn sync back on when ready

  • Re-save the files

  • This process is easier with version control e.g. Git set up, since you can see the changed files easily. Before taking the MCP out of alpha, we intend to improve the version control tools available in CLI and MCP combined, to give you the tools you need to make production work safer.

Troubleshooting

The assistant doesn’t have Siteglide tools
Enable the siteglide MCP server in your editor settings and reload.

Tools fail to start
Re-run siteglide-cli pull so MCP paths are correct for your machine.

Validation passes but the site looks wrong

• Check that sync is running and pointed at the right environment

• You can pass the URL of the page to the AI and it can visit and check for errors (may use more tokens)

• You can copy and paste error messages from page and console into the chat and it will attempt to fix.

Coming Soon

We’re working hard on our next CLI + MCP update which will be designed to be:

• Safer

• More collaborative

• More personalised

Watch this space!

Summary: Siteglide MCP makes your AI assistant a Siteglide expert!