CatchCMS exposes admin operations to AI clients through MCP (Model Context Protocol). With one endpoint and two tools, teams can query and maintain site data in a conversation, and every operation follows the permissions of the admin account that owns the token.
Step 1: Issue an MCP access token
A super administrator opens "Content Management → MCP Access Tokens", selects an enabled admin account, and creates a token with a name and a validity period (90 days by default, 1–365 days optional).
The full token is shown only once in the creation result and is cleared after the dialog closes. If it is lost, revoke the old token and create a new one. The token list shows the owner account, creation time, last-used time and expiry, and supports revocation. The connection stops working when the owner account is disabled or deleted, or when the token expires or is revoked.
Step 2: Connect from an AI client
In an MCP client that supports a custom Bearer token, set the endpoint and header:
- Transport: Streamable HTTP
- URL: https://your-domain/api/mcp
- Authorization: Bearer <the full token copied at creation>
To verify the connection, choose Streamable HTTP in the MCP Inspector, enter the same endpoint and header, and check the server information and tool list.
Step 3: Find and execute admin operations
Once connected, the client uses two tools: search finds operations by operationId, path or description and returns their input schema, then execute runs the selected operation. execute accepts operation_id, site_id, language, path, query and body.
For example, query sites and languages first, then create a category:
- search: {"query":"cmsCategoriesStore","limit":5}
- execute: {"operation_id":"cmsCategoriesStore","site_id":1,"body":{"model_id":1,"title":"News Center","type":1,"slug":"news"}}
Site-level operations require site_id; language is optional and selects the target language. For create, update and delete operations, query first to confirm the result and avoid duplicate writes.
Permissions and security boundaries
search returns only operations the current account is allowed to use, and execute re-checks permission on every call, including reads. The executor accepts only an operation ID and structured parameters; clients cannot specify URLs, headers, SQL or scripts. Write operations are recorded in the operation log. When AI hits a permission error, it should stop that operation and explain the reason.
Typical operations
- Maintain multi-site and multilingual content in batches
- Configure sites, categories, content models and dynamic fields
- Update themes, settings, forms and SEO metadata
- Query visit data, crawler data and content performance