Ghost MCP Server: manage Ghost CMS from Claude, Cursor and other AI clients
Ghost MCP Server is an open-source Model Context Protocol server for Ghost CMS, written in TypeScript. It connects an MCP client such as Claude Desktop, Claude Code or Cursor to the Ghost Admin API over stdio. With it, an AI assistant can create and edit posts and pages, set SEO metadata, manage tags, members and newsletters, upload images and themes, and export content.
It exposes 60 tools: 55 dedicated API tools, three shortcuts for SEO and tags, a route search and a generic request tool. The built-in catalog covers 253 Admin API routes taken from a pinned Ghost source snapshot. What works on your site depends on your credentials, your Ghost version and its feature flags. Routes and sources are listed in GHOST_API.list.

Quick start
You need Node.js 20 or newer and a Ghost site with Admin API access.
- In Ghost Admin, open Settings → Integrations → Add custom integration and copy the Admin API Key (
id:secret). Do not use the Content API key. - Add the server to your client configuration. For Claude Desktop or Cursor:
{
"mcpServers": {
"ghost": {
"command": "npx",
"args": ["-y", "@proars/ghost-mcp-server"],
"env": {
"GHOST_API_URL": "https://your-blog.com",
"GHOST_ADMIN_API_KEY": "64f1a2b3c4d5e6f7a8b9c0d1:3e5f7a9b1c2d4e6f8a0b2c4d6e8f0a1b3c5d7e9f1a2b4c6d8e0f2a3b5c7d9e1f"
}
}
}
}
The key above is a made-up example in the real format: 24 hex characters, a colon, then 64 hex characters. Replace both values with your own.
When you run the server with npx, pass the settings through env as shown. A .env file is read only from a source checkout (see below), or from the path in GHOST_MCP_ENV_FILE.
- Restart the client and ask something like "List my latest posts with their SEO titles."
GHOST_API_URL is the canonical Admin domain, including any site subdirectory, without /ghost/api/admin. For a local site, use an address such as http://localhost:2368.
Config file locations: Claude Desktop on macOS uses ~/Library/Application Support/Claude/claude_desktop_config.json. Cursor uses ~/.cursor/mcp.json or .cursor/mcp.json in your project.
Claude Code:
claude mcp add ghost --scope user \
-e GHOST_API_URL=https://your-blog.com \
-e GHOST_ADMIN_API_KEY=64f1a2b3c4d5e6f7a8b9c0d1:3e5f7a9b1c2d4e6f8a0b2c4d6e8f0a1b3c5d7e9f1a2b4c6d8e0f2a3b5c7d9e1f \
-- npx -y @proars/ghost-mcp-server
--scope user makes the server available in every folder. Without it, the server is registered only for the folder where you ran the command. Claude Code starts and stops the server itself, so you don't need to keep a terminal open.
Check the connection with claude mcp get ghost, which should report Connected, or type /mcp inside a claude session. To remove the server, run claude mcp remove ghost -s user.
Any other stdio MCP client can run the same command, npx -y @proars/ghost-mcp-server, with the two environment variables.
Updating: npx uses a cache, so @latest may not be checked against the registry on every start. To force a registry freshness check whenever the command runs, use npx --prefer-online -y @proars/ghost-mcp-server@latest. In a JSON client configuration, set args to:
["--prefer-online", "-y", "@proars/ghost-mcp-server@latest"]
To stay on one version, pin it, for example @proars/[email protected]. See the changelog for what changed in each release.
Don't run npx from inside a clone of this repository: npm finds the local package.json with the same name and fails with ghost-mcp-server: command not found. Run it from any other folder.
Install from source
Use this if you want to change the code or run an unpublished version.
git clone https://github.com/proars/ghost-mcp-server.git
cd ghost-mcp-server
npm install
npm run build
cp .env.example .env
Fill in GHOST_API_URL and GHOST_ADMIN_API_KEY in .env, then point the client at the build. Keep --silent: npm's startup banner would otherwise end up in the MCP protocol stream on stdout.
{
"mcpServers": {
"ghost": {
"command": "npm",
"args": ["--silent", "--prefix", "/ABSOLUTE/PATH/ghost-mcp-server", "start"]
}
}
}
The .env file in the project root is read regardless of the directory the client starts the server from. Variables that are already set take priority, and if both required variables are set, the file is not read at all. GHOST_MCP_ENV_FILE selects a different env file, and GHOST_MCP_SKIP_DOTENV=1 turns file loading off.
What can the server do?
| Resource | Tools |
|---|---|
| Posts | get_posts, get_post, get_post_by_slug, create_post, update_post, delete_post, copy_post |
| Pages | get_pages, get_page, get_page_by_slug, create_page, update_page, delete_page, copy_page |
| Tags | get_tags, get_tag, get_tag_by_slug, create_tag, update_tag, delete_tag |
| Tiers | get_tiers, get_tier, create_tier, update_tier |
| Newsletters | get_newsletters, get_newsletter, create_newsletter, update_newsletter |
| Offers | get_offers, get_offer, create_offer, update_offer |
| Members | get_members, get_member, create_member, update_member, delete_member |
| Labels | get_labels, create_label, update_label, delete_label |
| Staff | get_users, get_user, update_user, delete_user, get_roles, invite_user |
| Images and themes | upload_image, upload_theme, activate_theme |
| Webhooks | create_webhook, update_webhook, delete_webhook |
| Site | get_site, get_settings |
| Shortcuts | update_post_seo, update_page_seo, update_post_tags |
| Discovery and generic calls | ghost_list_endpoints, ghost_api_request |
Ghost has no delete endpoint for tiers, newsletters or offers, so those resources have no delete_* tool.
Warning: this server exposes every supported function, including destructive ones.
delete_post,delete_page,delete_tag,delete_member,delete_user,activate_theme,upload_themeandghost_api_requestcan permanently remove or overwrite content, subscribers and site settings. Deleted records usually cannot be restored. Create a separate integration key for the server, export a backup of your content from Ghost Admin before bulk work, and turn on manual approval for allupdate_*,delete_*and generic calls in your client.
How do I call the tools?
- Create and update: put the record fields in
dataand pass the recordidas a separate argument. Post and page updates needdata.updated_atfrom a fresh read. The SEO shortcuts (update_post_seo,update_page_seo) and the tag shortcut takeupdated_atdirectly. - Lists: use
limit,page,filter,order,includeandfields. Posts and pages come back with compact fields and tags by default. Single-record reads return the full record. - SEO:
update_post_seoandupdate_page_seosetmeta_title,meta_description,og_titleandog_descriptionin one call. Passid, the currentupdated_atand at least one of those fields. Fields you leave out stay unchanged. - Tags: setting tags replaces the whole set. Tags that don't exist yet are created, and an empty array removes all tags.
- Publishing: call
update_postwithdata.status="published", or with"scheduled"and a futuredata.published_at. To send the post by email, addnewsletter(a slug) and, optionally,email_segment. - Images:
upload_imagetakesfile_path, an optionalpurpose(image,profile_imageoricon) andref.
How do I reach routes that have no dedicated tool?
Use ghost_list_endpoints to search by group, method, access or search. Set without_dedicated_tool=true to list the other 198 routes. Each entry shows access requirements, feature flags and upload fields.
Call any route with ghost_api_request. Its path is relative to /ghost/api/admin, contains real IDs or slugs, and has no query string.
ghost_api_request method=GET path=/labels/slug/vip/
ghost_api_request method=POST path=/members/upload/ files={membersfile: "/Users/me/members.csv"}
ghost_api_request method=GET path=/posts/export/ save_to="/Users/me/posts.csv"
Send JSON payloads in body, using the structure Ghost expects. Put query parameters in query. Send multipart files as files={formField: "/absolute/path"}, with optional text fields.
For downloads, save_to must name a new file in an existing, writable folder. Trailing separators are rejected and existing files are never overwritten. When a POST, PUT or DELETE returns a file, set save_to on the first call, because repeating a mutation to get its response can repeat its effect.
Which Ghost permissions does it need?
| Class | Routes | Requirement |
|---|---|---|
I / PUBLIC | 125 | Integration-key allowlist or public access. Handler permissions still apply. |
U | 116 | A staff user's access token with the required role, or a user session. |
SESSION | 3 | User cookie session. Integration keys and staff tokens are blocked. |
FLOW | 9 | Login, recovery and setup flows with their own authentication. |
To use a staff token, put the token from the user's Ghost Admin profile into GHOST_ADMIN_API_KEY. Its permissions follow the user's role. Staff tokens cannot call DELETE /db/, PUT /users/owner/ or POST /authentication/reset/.
The server does not implement cookie sessions, login or 2FA. Those routes are in the catalog for discovery only.
Requests send Accept-Version: v5.0. The catalog is a snapshot of Ghost's main branch, so it includes newer or experimental routes that may return 404 on your installation. The public Content API is not included.
Is it safe to give an AI client access to my site?
Treat it like any account with write access to your blog, and keep a human in the loop.
- Keep keys in your client's environment or in
.env, and never commit them. Turn on manual approval for destructive tools: allupdate_*anddelete_*tools,activate_theme,upload_themeandghost_api_request. - If a mutation fails after it was sent, check Ghost before retrying, because it may have gone through. Use the read tools to inspect the result of a truncated response.
- Redirects are never followed. Configure the canonical Admin URL, and after a mutation receives a redirect, check the result before sending another request.
- Local file paths must be absolute. Hidden path components are refused, both as written and after symlink resolution. Uploads require regular files, and downloads are reserved before the request with owner-only permissions.
- The file checks are not a sandbox. They don't protect sensitive files that have visible names, and they don't stop a local process from rewriting directories during a call. Treat site content as untrusted and review upload requests.
- Uploads and downloads are buffered in memory. HTTP requests time out after 120 seconds, and tool responses longer than 100,000 characters are truncated.
Development
npm run typecheck # tsc --noEmit
npm test # regression tests with synthetic credentials and a local fake Ghost server
npm run catalog # regenerate src/catalog.ts from GHOST_API.list
npm run build # compile to dist/
npm start # run the build over stdio (requires configuration)
The route list lives in GHOST_API.list. scripts/generate-catalog.mjs turns it into src/catalog.ts, which you should never edit by hand. To change routes, edit the list, run npm run catalog, then run the tests. Logs go to stderr, because stdout is reserved for MCP messages.
GitHub Actions runs typecheck, build and tests on Node 20 and 22 for every push to main and every pull request.
The tests check catalog and tool consistency, request construction, file safety, redirects, retry guidance and configuration isolation. They don't prove compatibility with a live Ghost instance.
FAQ
Does it work with self-hosted Ghost and Ghost(Pro)?
Yes, as long as you can reach the Admin API and have a key. Set GHOST_API_URL to the API URL shown on your custom integration page in Ghost Admin. This matters most for Ghost(Pro) and for sites behind a custom domain, because the server never follows redirects: an address that redirects to another one will fail.
Which clients does it support? Any MCP client that can start a stdio server, including Claude Desktop, Claude Code and Cursor.
Can it read my public content without an Admin key? No. It uses the Admin API only. The public Content API is not covered.
Can it delete my content?
Yes, delete_post, delete_page and similar tools are permanent. Require manual approval for them in your client.
License
MIT.