PageForge checks every MCP request in layers, and each layer runs on every call:
- Credential. The request carries a valid OAuth token or personal access token in a header.
- Audience. An OAuth token must have been issued for this site’s MCP endpoint. A token minted for another server is refused.
- Scope. The tool being called must be covered by the scopes the connection was granted.
- WordPress permissions. The tool runs as the WordPress user who approved the connection, and that user’s own capabilities are checked.
- Tool checks. Each tool validates its own input and checks access to each post it touches.
Scopes
| Permission shown on screen | Scope | What it allows | Edition |
|---|---|---|---|
| Read PageForge data | pageforge:read | View templates, generated pages, SEO status and diagnostics | Free and Pro |
| Preview and validate | pageforge:preview | Run previews and validation without changing anything | Free and Pro |
| Create and update content | pageforge:write | Create, update and publish generated pages | Pro |
| Operate campaigns | pageforge:campaigns | Start, pause, resume, retry and cancel campaigns | Pro |
| Manage SEO | pageforge:seo | Update meta titles, descriptions, schema and internal links | Pro |
| Manage data sources | pageforge:content | Manage datasets, CSV imports and Google Sheets connections | Pro |
| Read analytics | pageforge:analytics | Read analytics data for generated pages | Pro |
| Administer PageForge | pageforge:admin | Change settings and connector configuration | Pro, never granted by default |
Free can only ever grant the first two. Pro scopes also need an active license, checked on every call, so a lapsed license stops write access at once. Broader scopes include the narrower read and preview scopes they depend on, but Administer PageForge never includes content changes, and content scopes never include administration.
When a call lacks a scope, PageForge answers with a precise challenge so the client can ask for exactly what it needs:
HTTP/1.1 403 Forbidden
WWW-Authenticate: Bearer error="insufficient_scope", scope="pageforge:seo", ...
Who can connect
By default only administrators (users who can manage options) can approve an AI client or create a token. If that account later loses the right, its tokens stop working immediately. A personal access token can never hold more than its owner could do: write scopes need the right to edit posts, and the administration scope needs administrator rights. In Pro, publishing through MCP also requires the right to publish that post type.
Other protections
- Rate limiting. Failed sign-in attempts are limited per IP address, and tools are metered by cost: reads are allowed more often than previews and changes. A client that hits a limit gets a 429 response with a
Retry-Afterheader. - Origin validation. Requests sent by a browser must come from your own site,
chatgpt.com,chat.openai.comorclaude.ai. Other origins are refused, which blocks DNS rebinding attacks. Server-to-server requests without an Origin header are allowed. - No caching. Authenticated responses are sent with
Cache-Control: no-store, so proxies and browsers do not keep copies. - Hashed secrets. Personal access tokens and OAuth client secrets are stored as keyed hashes.
- Short-lived tokens. OAuth access tokens expire quickly, and refresh tokens rotate on each use.
- Headers only. Tokens are read from headers, never from URLs.
Revoke access
| To stop | Do this on PageForge > Connections |
|---|---|
| One OAuth connection | Disconnect under OAuth grants |
| One token | Revoke under Personal access tokens |
| A suspected leaked token | Rotate it, then update the client |
| Every connector using a custom OAuth client | Revoke under OAuth clients |
| Everything | Untick the enable checkbox under MCP / AI Clients |
Tip Create one token per client and per machine, so you can revoke one without touching the others.
Next: Let assistants make changes and Cursor, VS Code, Claude Code and other clients.
