Hey Mike!
Great question, and one we discuss internally a lot when designing new KentiCopilot features.
The short answer for today: no autonomous writes in production, because the Management API and Management MCP server are preview features intended for local development only. We call this out explicitly in our docs as well: Enable the management API
The management API authenticates with a single shared secret, with no authorization options for different operation types. There is no user identity in the request, so there is nothing to scope permissions against and nothing to attribute a change to. So on two of the things you asked about, least privilege and an audit trail tied to the requester, we're not there yet.
Some guardrails and safety nets already exist today:
No publish operation. The MCP can create, read, update, delete, move and unpublish content items and pages. It cannot publish. Going live still requires a person in the admin UI.
Tool filtering. You can restrict the exposed toolset, and we do offer some recommended combinations here. Worth flagging that this is client side config, so it shapes what the agent reaches for rather than preventing anything at the API.
Versioning as a rollback layer. With content versioning enabled, every API call that modifies versioned content creates its own version entry. Verbose after a bulk run, but each step is individually revertible.
CI as the review gate. CI serialization is automatically disabled during management MCP requests, so you run CI store manually afterwards. In practice that means an agent's schema changes go through your normal PR review before they reach any other environment.
On where this goes next. While I can't share a specific timeline right now, we're looking into supporting content modelling further with agentic tooling, and extending the MCP scope to cover Forms. For both, we asked the same two questions: what's the intended workflow, and how do we make it safe when some actions are destructive.
We are considering a few approaches, and we'd love your feedback:
Destructive actions become fully reversible, for example soft deletes only via the tool
Block destructive actions when content is affected, for example you can't delete a form that has submissions
Instructions in the tooling itself so agents have to seek approval
Checks at a more critical point, for example confirming destructive changes before content sync runs between environments
Two things complicate all of that.
First, we don't control how a developer runs their agent. In yolo mode, soft gates could get bypassed and based on Anthropic's findings, developers also often approve destructive actions when they're in the flow.
Second, there is an underlying API, and guardrails meant specifically for agents are not going to necessarily work. E.g. we might block deletions via the MCP tool, but an agent could totally send a DELETE/PATCH request to the API directly. Strict rules would have to live on the API layer rather than the AI tooling itself.
So it's likely a combination: real safety nets like rollback and undo, plus a balance between tooling restrictions and unsupervised capability. Possibly a per-project decision. Relaxed for autonomous local dev, strict once a project is anywhere near production.
On your specific question about staging: nobody should be running this against staging today given the auth model. If you can describe what you'd want to run there, and what approval and audit you'd need to be comfortable with it, that's exactly the information we need to support you 😀