In the span of a few days, three CMS vendors opened official Model Context Protocol paths that let AI coding agents work on their platforms. Last week, Sitecore released its Developer MCP server in beta. A day earlier, Webflow launched an MCP connector for Microsoft Copilot. Earlier in the month, Sanity added Media Library tools to its hosted MCP server.
Those are vendor announcements, and we have not independently tested the products. The coverage we found is largely restatements of those announcements and official docs, not hands-on reviews. What the docs do make clear is that the agents can already do more than read. They can create components, collections, pages and assets. The more useful question, then, is not whether MCP has arrived in the CMS, but which writes are gated, which ones apply immediately, and what a DevOps team should lock down before anyone turns a connector on.
What Developers Can Do Today
Sitecore's Developer MCP server is aimed at developers and AI agents working in tools such as Cursor, Visual Studio Code, Claude Code, Codex and OpenCode. Official docs describe 13 tools for component management, content modeling, site discovery and content-item creation under an existing parent, using a specified content type. The production server uses Streamable HTTP at a Sitecore-hosted URL, with the server name sc-developer. Sitecore says the server is in beta and is seeking feedback as it works toward general availability. A sample prompt in the docs walks an agent through creating a Testimonial component and content type, then reminds teams that the steps can vary by model, prompt and site, and that they should review all changes before they publish.
Webflow is opening a different front door. The company's recent announcement says the Microsoft Copilot MCP Connector can create pages, manage and publish CMS content, pull Analyze and Optimize insights, and audit a site for AEO, SEO, accessibility, broken links and content quality. Webflow says those are the same core capabilities already available through Claude, ChatGPT, Codex and Slack, now with Copilot Skills that follow Webflow's own practices for jobs such as collection setup, accessibility audits and pre-publish checks. Official help documentation describes the connector as free to install and says it does not consume Webflow AI credits, though capabilities still depend on a user's plan and permissions.
Sanity's November releases added Media Library tools to the company's existing hosted MCP server, when Media Library is enabled for the organization. Official docs say a connected agent in Claude Code, Cursor or another MCP client can upload images, videos and files from a public URL or from a local file up to 50 MiB, search by aspect, type, folder or title, edit titles and aspect values, organize collections, link assets to documents, and publish or discard drafts. The agent can also list the documents that use an asset. Those are Sanity's descriptions of the current tool set, and they go a little further than the November 9 blog post, which focused on upload, search, draft edits and visibility. We have not verified the tools in a live library.
This is the developer path, not the marketer one we covered earlier this month. Sitecore's MarketingIQ turns gaps in AI answers into staged content work a person still has to approve. Adobe put CX Enterprise Coworker in GenStudio, also with MCP in the mix and a human still making the call. The news this week is that coding agents can reach the CMS through the same open protocol, from the IDE or from Copilot, and create work that may need the same kind of review.
CrafterCMS, for its part, lists an official MCP server plugin that exposex content, services and tools over HTTPS, with secured and public modes and OIDC or OAuth single sign-on. How Access Is Governed
The three vendors do not govern writes the same way.
Sitecore authenticates the Developer MCP server with MCP OAuth and no custom headers. The client opens a browser sign-in prompt, then talks to the hosted server. Official docs tell teams to review all changes before they publish. They do not, in the pages we read, describe a confirm-before-write prompt at the tool level, or an activity log that attributes work to an agent. That review step is documented guidance. We have not confirmed how tightly the product enforces it in beta.
Webflow is more specific about the confirmation gate. Official help says Copilot reads from Webflow without asking, and that, by default, it asks the user to confirm each time before it creates, updates or deletes anything. The user can allow the action once, for the rest of the conversation, or always, and each choice applies to the individual tool, not to the whole connector. Access follows existing Webflow roles and permissions. Any site role except Reviewer can use the connector, and the user chooses which sites or workspaces to authorize. Webflow's announcement also says work appears in activity logs with AI attribution. The license bar is on the Microsoft side: each person needs a Microsoft 365 Copilot add-on or Microsoft 365 E7. The app listing notes that federated Copilot connectors are not supported with Copilot Studio licenses or Microsoft 365 Copilot pay-as-you-go.
Sanity runs the Media Library tools as the authenticated account, or as an API token for an account with the same access. The agent can only see or change what that person can see or change in the Media Library app, and the account has to belong to the organization that owns the library. Every write is validated against the library's schema before anything is stored. Title and aspect changes are saved as drafts, which a person can review, publish or discard. Visibility changes, and collection creates and edits, apply immediately and do not create drafts. Official docs also say the agent cannot delete assets or collections, cannot change aspect definitions, and cannot place assets in folders.
Those differences matter more than the protocol they share. MCP standardizes how an agent finds and calls tools. It does not, by itself, decide whether a write is a draft, a confirmation prompt, or a live change.
What Teams Should Lock Down First
If you are going to enable these connectors, treat them as production access, not as a convenience in the IDE.
First, inventory who can install an MCP client on a developer machine, and under whose identity the agent will sign in. Sitecore's OAuth flow and Sanity's signed-in account model both inherit whatever that person can already do. A principal with broad authoring rights becomes a principal that an agent can exercise at typing speed.
Second, keep write confirmations on until you have a reason to turn them off. Webflow's "always allow" setting is per tool, and it is easy to widen by accident. Sitecore's review-before-publish guidance only helps if someone actually reviews the work. For Sanity, remember that visibility and collection changes are not drafts.
Third, scope the connection. Webflow lets a user authorize specific sites and workspaces. Use that. Do not connect an entire company workspace because a single marketing site needed an accessibility audit. On Sitecore, the beta server can create components, content types and items; decide whether that belongs in a production tenant or in a lower environment while the feature is still in beta.
Fourth, watch the audit trail. Webflow documents AI attribution in activity logs. Ask Sitecore and Sanity, in your own tenants, how you will tell a person's edit from an agent's. If you cannot answer that, you are not ready to let the tools publish.
Fifth, match the license to the product you think you bought. Webflow's connector does not consume Webflow AI credits, but it does require a qualifying Microsoft 365 Copilot license, not a Copilot Studio or pay-as-you-go path. Sitecore's Developer MCP is a beta, not general availability. Sanity's Media Library tools only appear if Media Library is enabled and the account can reach it.
The protocol is the easy part. The harder work is the same as it is for any new API into the CMS: least-privilege identities, a publish path that still requires a person, and a log you can actually read when something ships that shouldn't have.
Suresh Venkat