“Write a better headline” asks for a suggestion. “Change the headline in the draft” asks for an intervention. “Publish the new version” changes the page for visitors. The three sentences are close together in wording and describe very different end states in technical terms.
Framer announced its MCP integration on 22 October 2026. According to the original announcement, connected AI tools can work on the design canvas, manage CMS content and publish. The news makes a general shift visible: a conversation can turn into work on the actual project. Which action is permitted in a specific assignment, however, is not decided by the mere existence of the connection.
Model, client and tool server are different roles
The model processes the context and, among other things, proposes tool calls. The AI application in use connects these tools via a client. The server describes and handles the functions on offer. The official MCP architecture separates these roles; the standard does not specify how each AI application runs its model internally.
This distinction helps with your own or local models. A model name alone does not prove a working connection. The client in use has to support the necessary transport and authentication. The specific server then has to offer exactly the tools that the task requires.
For the event example, this means: an existing connection can open up technical possibilities. It does not yet say whether the client selects the right draft, whether the texts have been confirmed or whether publishing is part of the assignment.
A change request needs an unambiguous object
A usable instruction names the page, the permitted changes and the end state. For example: “In the specified draft, change only the headline and introduction. Keep the date and the condition of participation unchanged. Do not publish anything. Afterwards, list the places affected.” This template of our own does not claim any special Framer safeguard.
Before the change, the current state of the project should be read. Otherwise an instruction may be based on an outdated version of the text. After the change, the feedback needs a reference to the starting point: what was replaced, what was kept and where is the result? A nice paragraph in the chat is not enough for that.
A documented remote MCP route at KENJTEC
The KENJTEC MCP guide describes such access for suitable clients. In KENJTEC’s documented test bench, Claude Code is recorded as a connected program; this does not imply a passed run with every other client. The tools on offer are filtered according to the permissions granted.
sites:read → list_sites, get_site
sites:write → create_site, edit_site,
publish_site, unpublish_siteAmong other things, get_site returns the ID, page overview and draft status, not a complete source tree that can be edited at will. Creating and changing a website runs through Kenji and uses KENJTEC AI credit. Your own external AI issues the instruction; it does not automatically write its complete website code straight into the site.
The scope of the write permission is important: it also covers publishing. Labelling tools that have significant consequences helps a client to ask for confirmation, but it is not an additional server-side confirmation for each individual call. An instruction such as “draft only” has to be actually respected by the client. For pure advice, you can stay with read permission and apply the selected suggestion yourself in the editor.
An open standard means an understandable contract
Even access that works technically is not root, SSH or general storage access. For a specific connection that has been set up, what counts are the tools on offer, the permissions chosen and the rules of the application concerned. Framer and KENJTEC offer different sets of functions; the same protocol name does not make them interchangeable.
The practical next step is therefore small: name a test project of your own, read its current state and define a limited desired end state. Only then does a change make sense. If, after the run, you can explain which tool changed what, and why nothing is public yet or something already is, the AI connection becomes a traceable workflow.
Sources and status
Research and documentation as of 25 October 2026. Framer MCP · Official Framer updates overview: Framer MCP, 22 October 2026 · Architecture overview.
