Introducing the new DatoCMS Remote MCP
Now, as a marketer myself, I can definitely tell you that I don't use AI to create my content and I'm keeping it natural. Insert wink wink meme. If you're the same as me, then you've probably been doing a lot of copy-pasting from the content you've been manually creating in ChatGPT or Claude or wherever else. Prompt something in Claude, copy it into DatoCMS, edit it. Ask ChatGPT to translate something, paste it back in. Do this for every new record but split the copy paste operations into as many fields as that model has. I mean. It gets the job done, but it's a bit silly when you think about it. I hate how much my cmd, c, and v keys are wearing out.
The new DatoCMS MCP server fixes that. It's a secure connection between your AI assistant of choice and your DatoCMS project, so instead of pingponging content back and forth between tabs, you just describe what you need and the AI handles it directly.
Never heard of an MCP before? Don't worry, we'll quickly look into that.
Heard of an MCP before? Check out our user guide on using the DatoCMS MCP for content.
And if you're a dev who's going to primarily work with the CLI, the MCP is still really cool and all, but you should totally check out working with DatoCMS Skills for that lil extra something something.
What's an MCP
We're not looking to be the SEO authority on What is an MCP, so if you're keen to dive in to details, check out the Model Context Protocol docs. It's basically a standard way for AI assistants to connect to external tools and actually do things in/with/through them, and not just talk about them.
Without an MCP, your AI assistant is like a very smart colleague who can only tell you what to do in X tool. With an MCP, it's like giving them a comfy little chair and having them do it for you.
Once it's connected, you can ask your AI assistant to do things in DatoCMS in plain language, and it will do that on your behalf (note: on your behalf means AS YOU. Your account is what will be used by your AI).
What you can actually do with it
About anything you'd normally do by clicking/typing around in DatoCMS, you can now just ask your AI assistant to do instead:
"Create a new blog post with this title and body content and ensure all validations are met."
"Add German and Italian translations to all landing pages published in April."
"Unpublish all the published records, I'm not happy with them."
"Find all records that are missing an SEO description and list them for me. And for all the records with an SEO description, audit them and improve them for me to review before publishing."
"Update the author field on these 15 records to link to the new author profile."
"Under Product Categories create a new category for socks, and then create 50 dummy products with unique images before localizing them into Mandarin and trigger the translation workflow for me."
"Ooh I forgot to add images. Install the Unsplash Asset Source plugin for me, and add 10 images of shawarmas for me to look at and connect to the posts."
It handles long records, lots of fields, deeply nested content structures, localizations, constraints, etc. The previous version of our MCP (yeah, we had one) could struggle with complex content. This one doesn't.
You're in control of what it can touch
When you connect your AI assistant to DatoCMS via the MCP, you can choose exactly which projects it can access, and it only ever has the same permissions your DatoCMS account has. It can't access anything you can't access yourself.
On top of that, read operations (things like looking things up, searching, listing records) can run freely without any extra confirmation. Write operations, which is anything that creates, updates, or deletes content, will require your explicit approval before they run.
So the assistant will tell you what it's about to do and wait for your OK. Nothing happens behind your back.
How to connect it
The whole point of making this remote is that there's nothing to install. No terminal, no dev environment, no asking your developer to set something up.
For Claude (web or desktop):
Go to Customize, then Connectors, click the plus button, select Add custom connector, enter DatoCMS as the name and https://mcp.datocms.com as the URL, and click Add. Then hit Connect and log in with your DatoCMS account.
For ChatGPT:
Enable Developer Mode under Settings, then create a new app with the MCP Server URL set to https://mcp.datocms.com. Connect and authenticate.
Both work in the browser and on mobile. Full setup instructions for every supported client are in the docs.
One script, not fifty tool calls
Most MCPs work by exposing API endpoints directly. When you ask your fave LLM to do something, it makes a tool call, gets a result, makes another tool call, gets another result, and on and on. For anything big, this gets expensive fast (in tokens and in round-trips and in time) and the agent tends to lose context and coherence across a long chain of individual calls.
Our MCP does things a little bit differently. When you give it a task, it writes a TypeScript script that batches the entire operation (multiple API calls, conditional logic, whatever is required) and executes it in an isolated sandbox through one execution, not fifty shades of tool calls.
The practical result is that operations that would have been slow or unreliable before, like long records, deeply nested blocks, batch localizations, or bulk updates across many records, now complete fast and accurately.
Read vs. write: two tools kept separate
The MCP has two execution tools and they behave differently by design.
Read-only operations use upsert_and_execute_safe_script which run against a read-only API token and most clients will let them through without a confirmation prompt.
Write operations (and anything destructive) uses upsert_and_execute_unsafe_script which has full read-write permissions, and the client asks for your explicit confirmation before running. So a setup where reads are always allowed and writes require approval is just the default, with nothing to configure on your end.
Beta note
This is a beta. It works well, but there are usage limits depending on your DatoCMS plan, and we may adjust things as we learn more from real-world usage. If something doesn't behave as expected, we genuinely want to hear about it.
PS: If you're a developer, check out the new Agent Skills we're launching alongside this.
