Connect to Creght over MCP

Creght runs an MCP server. Once an AI assistant (Claude, Cursor and others) is connected, it can use your account to list sites, check publish status and visit analytics, read and write CMS content, read form submissions, upload assets and publish sites. Authorization is OAuth in the browser; there is no key to copy.

Endpoint

Your account is on	MCP URL
creght.com	https://creght.com/api/mcp
creght.cn	https://creght.cn/api/mcp

• The path is /api/mcp, not /mcp. https://creght.com/mcp returns the editor page, and the client reports an invalid MCP response.
• Accounts and data on the two sites are separate: an account registered on creght.com only works with the creght.com URL.
• The transport is Streamable HTTP (stateless, one request per call). The legacy SSE transport is not supported.

Connect a client

The examples use creght.com; swap in the creght.cn URL if that is where your account is.

Claude Code

claude mcp add --transport http creght https://creght.com/api/mcp

Then run /mcp in Claude Code and pick creght to authorize.

Cursor and other JSON-configured clients

Add this to ~/.cursor/mcp.json (or .cursor/mcp.json in a project):

{
  "mcpServers": {
    "creght": { "url": "https://creght.com/api/mcp" }
  }
}

Any client that supports remote MCP with OAuth takes the same URL. There is no client id, secret or token to fill in.

Claude on the web and desktop

Settings → Connectors → Add custom connector. Set the URL to https://creght.com/api/mcp and leave the OAuth fields under advanced settings empty.

Authorization and permissions

The first call gets a 401, and the client opens your browser: sign in to Creght, approve on the consent page, and you are sent back to the client. The client handles the whole exchange; you only click Allow.

• The app name on the consent page is whatever the client reports, and it is marked unverified. The page also shows the callback address. Don't approve if that address isn't the client you are using (for example local 127.0.0.1, or claude.ai).
• There are two scopes: site:read (every read tool) and site:write (writing CMS content, uploading assets, publishing, managing webhooks). A client that requests no scope gets site:read only, and write tools then fail with a permission error.
• The grant never exceeds your account: you see your own projects and projects you were invited to, with the same read/write rights as your role in each project.
• Access tokens last 1 hour and the client renews them with a refresh token. The refresh token lasts 30 days and the clock restarts every time it is used, so a client you use regularly doesn't need to be re-authorized.

Tools

Every tool takes a project_id (site tools also take a site_id); get them from list_sites first.

Sites

Tool	Scope	Parameters	Returns / notes
list_sites	read	none	sites[]: project_id, project_name, site_id, site_name
site_status	read	project_id, site_id	Whether it is published, the live version, live_url (custom domain first; empty if never published), preview_url, and has_changes (whether the workspace differs from the live version)
site_publish	write	project_id, site_id, note (optional, kept in version history)	Publishes what the preview shows, same as creght publish. Errors if nothing changed. Returns the new status plus version_id and version_no
visit_stats	read	project_id, site_id, start_at / end_at (Unix seconds, optional), limit (top N per dimension, default 20)	summary (pv / uv / ip), trend (by UTC date), breakdowns (country, city, device, browser, OS, referrer host, channel, page). Defaults to the last 30 days; a start beyond your plan's range is narrowed, so trust the returned start_at

CMS

Tool	Scope	Parameters	Returns / notes
cms_collections	read	project_id	collections[]: id, key, name, desc
cms_collection	read	project_id, collection (key or id)	The collection and its field definitions fields (JSON Schema). Read it before writing content
cms_content_list	read	project_id, collection, limit (default 20, max 100), offset, status (online / offline), search	total, has_more, list[]: id, slug, status, body, created_at, updated_at
cms_content_create	write	project_id, collection, slug (unique in the collection), body	The new entry's id. Live as soon as it is created
cms_content_update	write	project_id, collection, id, slug (optional), body (optional)	body is merged into the existing entry; fields you don't send stay as they are. updated: false means the values were already the same. Changes go live immediately

Forms

Tool	Scope	Parameters	Returns / notes
form_list	read	project_id	forms[]: id, key, name, desc, fields, submissions_total
form_submissions	read	project_id, form_id (optional, id or key), since (RFC3339, optional), cursor, limit (default 50, max 200)	Returned in submission order with a next_cursor. For incremental sync, pass the last next_cursor back; nothing is repeated or skipped. The submitter's IP is not returned, only the country code resolved from it
form_webhook_create / list / delete / test	list read, the rest write	Push every submission to your https URL in real time; see Form submission webhooks

Assets

Tool	Scope	Parameters	Returns / notes
asset_upload	write	project_id, site_id, filename (with extension), content_base64 (a data: URL also works), content_type (optional)	A public url you can use directly in pages and CMS content. Same as creght upload; uploading the same file again reuses it. 20 MB per file (before base64)

A typical flow

Publishing a blog post:

1. list_sites to find the site's project_id and site_id.
2. cms_collections to find the blog collection, then cms_collection for its fields.
3. If there is a cover image, asset_upload it first and put the returned url in the right field.
4. cms_content_create. The post is now on the live list and detail pages; you do not need site_publish.

Use site_publish only after site code has changed: it publishes the page code and configuration in the workspace (what the preview URL shows). CMS content doesn't go through it.

Limits

• There are no CMS drafts. cms_content_create and cms_content_update both go live directly. If someone needs to review first, draft locally and call the tool after they approve.
• site_publish takes effect for every visitor at once. Have someone check the preview_url before calling it.
• The keys and types in body must match the fields returned by cms_collection.
• There are no tools to delete CMS entries, change a collection's structure or edit site code. Use the editor or the creght CLI.
• asset_upload carries the whole file in one call and rejects anything over 20 MB. Use creght upload for larger files.

Troubleshooting

Symptom	Cause
The client says the response isn't JSON or isn't a valid MCP response	The URL ends in /mcp; change it to /api/mcp
You can't sign in on the consent page, or it says the account doesn't exist	Wrong site: a creght.com account connected to the creght.cn URL (or the other way round). Fix the URL, then remove the server and add it again
Read tools work, write tools fail with a permission error	The client only requested site:read, or your role in that project is read-only
A site is missing	The project isn't yours and you weren't invited to it. list_sites is everything you can act on
site_publish says there are no changes	The workspace matches the live version; when site_status shows has_changes false, there is nothing to publish
asset_upload says the request is too large	The file is over 20 MB; use creght upload