Your content, from the chat: Contentrain Studio MCP Cloud is now in Claude's connectors directory

As of this week, Contentrain Studio MCP Cloud is listed in Claude's connectors directory. Open the directory in Claude, search for Contentrain, click Connect, pick a project in the browser, and Claude can read and edit that project's content. It is listed as a community connector: it went through the directory's automated review.

The listing's one-liner is "Edit your Git-backed product content, safely". This post is about that word "safely": what MCP Cloud can do, what it deliberately can't, and why there is a hosted endpoint at all when the Contentrain MCP server is MIT-licensed and runs on your laptop.

What MCP Cloud is

Contentrain keeps content as typed models in a Git repository: plain JSON and Markdown under .contentrain/, read by your site through a generated, typed client. Contentrain Studio is the web app on top of that repository: an AI- and Git-native CMS with roles, review and publishing.

MCP Cloud is Studio's remote MCP endpoint:

https://studio.contentrain.io/api/mcp/remote

It is one URL for everyone. Which project a connection works on is decided when you sign in, not by the URL. Authentication is OAuth 2.1 with PKCE, so Claude never sees a password or a GitHub token. The directory listing points at exactly this endpoint.

Connecting it: four steps, once per project

1. Find it. In Claude, open the connectors directory and search for Contentrain.
2. Connect. Claude discovers Studio's OAuth server on its own and opens Studio in your browser.
3. Consent. Sign in to Studio, choose the workspace and project this connection may use, and approve the permissions it asks for.
4. Use it from the chat. Ask "What content models and locales does this project have?" or "Which entries are still missing a German translation?".

A few things worth knowing up front:

• One connection is one project. To work on a different project, connect again and pick it. The old connection stays until you disconnect it.
• You stay signed in. Access tokens expire after an hour, and Claude refreshes them silently for up to 30 days.
• Media permissions show up only where they can work. The consent page lists media access only once you pick a project that can use media (more on that below). If a project becomes eligible later, reconnect it once.
• You can cut it off any time. Studio's Connected Apps settings list every connection with its client, project, permissions and monthly call count. Disconnecting revokes it immediately; the next request just asks the app to authorize again.

Not in Claude? The same endpoint works in other MCP clients. Studio documents setup for Claude Code, ChatGPT (developer mode), Codex and MCP Inspector:

claude mcp add --transport http contentrain https://studio.contentrain.io/api/mcp/remote

What Claude can do with it

On a connected project Claude gets 12 tools, or 17 on a project that can use media. They are the contentrain_* tools from the open MCP package, served by Studio:

Every write is validated against your schema before anything is committed, so the model can't invent a field or break a locale. Read tools and destructive tools are annotated separately, so Claude knows which calls change something. Deleting content needs an explicit confirmation.

Where the writes go

This is the part that makes it "safe" rather than just "convenient". Claude never writes to your main branch directly:

1. Each write lands as a Git commit on its own cr/* branch in your repository.
2. Studio then decides it exactly as it decides a write from its own editor:
   • on an auto-merge project, the change lands in seconds;
   • on a review project, the approval policy decides. The default policy holds every write for a person to approve in Studio.
3. Merging and deploying are not tools on this surface. If you ask Claude to "merge my changes and deploy the site", the right answer is that the project's workflow already handles the merge, and deployment belongs to your hosting.

The connection also acts as you, with your current role. A project editor can save drafts, but only workspace owners and admins can publish. The role is checked on every call, so someone removed or demoted after connecting loses the access their new role doesn't allow.

If you already use Studio: why connect Claude at all?

You already have a web editor. Here is what MCP Cloud adds to it:

• Your team works from the chat they already have open. "Shorten the excerpt of this post to one sentence", "which pages are missing a German version?", "add these three FAQ entries". Nobody opens a terminal or clones the repository.
• The same rules as the UI. Roles, review, the approval policy and publishing rights are Studio's, not a second set of rules for AI. An agent can't do anything the person who connected it couldn't do in Studio.
• Bulk and tedious jobs. Translating a batch of entries, filling a new field across a collection, or checking a project with validate before a release are all jobs a model handles well. Each change still arrives as a reviewable commit.
• Media from a URL. On projects with media, Claude can bring an image in from a URL into Studio's media library and set it on an entry. MCP has no channel for binary files, so uploads from your disk stay in the Studio UI.
• One audit trail. Whether a change came from the editor, Studio's own chat, or Claude through MCP Cloud, it is a Git commit in your repository.

If you don't use Studio: why not just run the local MCP?

You can, and for a developer working alone in a repository it is often the right choice. The Contentrain MCP server is MIT-licensed, needs no account, and runs over stdio next to your code with 23 tools, including the local ones the remote endpoint never offers (init, scaffold, scan, bulk, branch management).

The local server needs a checkout, Node and Git on the machine of whoever is editing, and an agent that runs next to that checkout. That's fine for you. It doesn't work for:

• People who don't have the repository. A marketer, a translator or a founder can connect Claude in a browser and work on content without cloning anything or knowing what a branch is.
• A team. OAuth means each person connects as themselves, with their own role, and an admin can see and revoke every connection in the workspace. A shared local setup has no such layer.
• Review before it ships. Locally, review means you push branches and someone merges them. In Studio, the review flow and approval policy are built in, and MCP Cloud writes go through them automatically.
• Media and delivery. The local server has no media tools. On eligible plans, Studio stores assets, serves them through its CDN, and exposes them to Claude.
• Chat apps without your repository. Claude on the web or on your phone can't start a local server next to your code. A remote endpoint is how those apps reach your content.

Both write the same .contentrain/ format into the same kind of repository, so you can start local and add Studio later, or use both on one project.

What it needs, and the honest limits

• A Studio account and project connected to a GitHub repository. MCP Cloud works on Studio's Starter ($9/month) and Pro ($49/month) plans, both with a 14-day trial, and on Enterprise. It also works on a self-hosted Community deployment that runs Studio's own sign-in with PostgreSQL. It does not work on the Free workspace.
• Calls are metered. Calls from Claude and from Studio's per-project MCP API keys share one MCP Cloud quota each month: 5,000 calls on Starter, 150,000 on Pro. Each connection is also limited to 60 requests a minute.
• Media needs an eligible project: a plan with media upload, CDN delivery turned on for the project, and Studio's media storage. Without those, the five media tools don't appear at all.
• No merge, deploy or branch management from the chat. Writes follow the project's workflow; publishing is for owners and admins.
• Claude's side has its own rules. Whether you can add connectors depends on your Claude plan, and on Team and Enterprise plans your admin decides which connectors are allowed.

Wrapping up

The local MCP server gave agents a safe way to edit content in a repository they can reach. MCP Cloud does the same for everyone else: people who live in a chat window, teams that need roles and review, and content that includes images. In both cases the result is the same, plain files and Git commits in your repository.

If you already have a Studio project, find Contentrain in Claude's connectors directory and connect it. If you don't, you can start a 14-day trial, connect a repository, and try the same prompts. The MCP Cloud docs cover the other clients, and the privacy addendum explains what MCP Cloud stores.

We'd like to know what you'd ask your content first. The comments are open.