What is an agentic CMS?
An agentic CMS is a content management system that AI agents can operate end to end: create the content model, write and update entries, upload media and publish, through the same operations and permission checks as the people on the team.
By the HelmCMS team · Reviewed against the HelmCMS codebase on October 29, 2026
What makes a CMS agentic
Most content systems now have some AI inside: a button that drafts a paragraph or suggests a title. That is an AI feature in a CMS. An agentic CMS goes the other way round: the agent sits outside, in the assistant you already use, and drives the CMS itself. For that to work safely, five things have to be true.
| Requirement | Why it matters | How HelmCMS does it |
|---|---|---|
| Every operation is a tool | If the agent can only write text, a person still has to build models, upload media and publish. | 42 MCP tools cover sites, content models, entries, versions, media, search and webhooks. |
| The schema is data, not code | An agent cannot open a pull request and wait for a deploy every time a page needs a new field. | Content models are created and changed at runtime, from the admin or over MCP. |
| Scoped sign-in | An agent should get exactly the access you choose, and you should be able to take it back. | OAuth 2.1 with four scopes: read, write, publish and admin. A token can never exceed the user's own role. |
| Drafting is separate from publishing | Letting an agent write is low risk; letting it change the live site is a decision. | Writing drafts and publishing are separate permissions, from the author role up to the editor role. |
| A record of what changed | When something goes wrong, you need to see who did it and roll it back. | Every save is an immutable version you can restore, and write and publish actions go to an audit log. |
How an agent builds a page in HelmCMS
Here is the sequence of tool calls an assistant makes when you ask it to set up a blog post type and publish a first post. The tool names are the real ones; the order is a typical one.
list_sitesfinds the site you mean.create_content_typeadds a “Blog post” collection with title, slug, rich text, cover image and author fields.import_media_from_urlbrings in the cover image (public URLs only; private and local addresses are refused).create_entrysaves the post as a draft, which creates version 1.publish_entrymakes it live, but only if the token was granted the publish scope and the user is an editor or above.
Your frontend then reads the post from the delivery API, and a signed webhook tells it to refresh.
The full tool surface
| Area | Tools | Examples |
|---|---|---|
| Organizations, workspaces, sites | 14 | create_site, get_site_overview |
| Content models | 5 | create_content_type, update_content_type |
| Entries and versions | 10 | update_entry, restore_version, move_entry |
| Media | 7 | create_media_upload, import_media_from_url |
| Webhooks | 5 | create_webhook, list_webhook_deliveries |
| Search | 1 | search_content |
Which AI assistants can connect
HelmCMS runs a remote MCP server over Streamable HTTP with OAuth sign-in. Any MCP client that supports remote servers with OAuth can connect, for example Claude, ChatGPT and Cursor. You paste one URL, sign in, and choose which scopes to grant. Clients that only accept a static token are not supported today.
Keeping people in charge
The simplest safe setup is to sign the agent in as a user with the author role: it can create and edit drafts and upload media, but it cannot publish, delete entries or change the content model. An editor reviews the draft in the admin, checks its version history and publishes.
A dedicated review queue, where every agent change waits for approval with a before-and-after view, is on the roadmap. Until it ships, the role split above is how teams keep the live site under human control.