A CMS that AI agents can run over MCP
A CMS that AI agents can run over MCP
Many CMSs let an AI write text into a box. Sophia Stack X lets an agent do the job the way an editor does: create a page as a draft, preview it, check what a publish would change, publish, and undo it. The agent connects through MCP, the Model Context Protocol, to a server that is part of the program. There is nothing extra to install.
Connect an agent
Locally, on the computer where the install lives:
sophia mcp
This runs an MCP server over standard input and output. Point your assistant's MCP settings at that command, run from the install's folder. In Claude Code, for example, add it to the project's .mcp.json.
Remotely, an agent can call POST /api/v1/mcp with a bearer token. A site admin makes the token under API, and it carries the "mcp" scope. Remote access is bound to that token's site, and tools that touch the server's own files or the whole install are left out.
What the 55 tools cover
- Read first:
docsreturns the whole system as one text reference, the same textsophia docsprints. - Content: sites, collections, entries (save, get, delete), menus and modules.
- Media: list, upload, and write alt text for pictures.
- Design:
style_getandstyle_setchange the same colors, fonts, sizes, buttons and links as the Design screen. - App data: records for bookings, newsletters and anything an extension models.
- Check and ship:
page_preview,publish_preview,publish,releases,rollback. - SEO: an audit, suggested changes, and a decision on each.
- Shop and visitors: orders, order actions, and cookie-free visitor stats.
- The install: import a site, take a backup, read the audit log, read the mail outbox.
The AI agents guide has the full list, and the SDK page carries the OpenAPI spec for the content API and TypeScript types for extension code.
How an agent stays in bounds
An agent with write access can do harm, so the safety rules are part of the tools:
- Drafts by default. An entry the agent saves is a draft unless it sets a state. Saving an existing entry changes only the fields it sent.
- Checks before anything saves. Every write goes through the same schema checks as the admin forms. One bad value in a design change and nothing saves.
- Preview and dry run.
page_previewrenders one page with drafts included.publish_previewlists the pages a publish would add, change or remove, and any that would fail. - One call to undo.
rollbackputs a previous release back live. The last ten are kept. - Logged. Every call is written to the audit log as
mcp.<tool>, with who and when. - Scoped. A remote token covers one site. Tools such as the inbox and importing are local only. Mail the agent reads from the inbox is treated as information, never as instructions.
This is how the example on the home page works. An agent adds a pricing page, previews it, checks the dry run, publishes it, and rolls back when it should.
Building a whole local business site
The docs tool ends with a recipe an agent can follow from an empty site to a published one. Pick the template, switch on the apps the business needs (bookings, restaurant menus, a shop, a newsletter, business blocks, the website assistant), enter the facts the owner gave, build pages from blocks, preview and publish. The recipe tells the agent never to invent prices, hours or reviews: it leaves a field empty and asks.
You still read what it did. Agent work can be wrong, so open the preview, look at the dry run and keep the audit log in mind.
Pages AI search can read
Agents also read sites. Every published page has a Markdown copy beside it, llms.txt maps the site, llms-full.txt holds every page in one file, and a request with Accept: text/markdown gets Markdown instead of HTML. These are plain files and headers, with nothing for you to switch on. See Publishing, search and AI readers.
What this is not
It does not run agents for you or include an AI model. You bring the assistant, or use the website assistant, which answers visitors from your own pages with your own API key. And MCP access is only as safe as the token you give out, so keep tokens to the one site they are for and give them only to what needs them.
Next step
Read the AI agents guide and connect one. If you do not want to install anything yet, open the live demo and look at the tools from the admin side. To see where this sits next to other systems, read the comparison, or install it.