Enables browsing and editing of Markdown documents through RefMD's API, with tools for listing, searching, creating, reading, and updating document content and metadata.
RefMD MCP Server
A Model Context Protocol server that exposes RefMD documents over a hosted SSE endpoint, so chatbots can browse and edit Markdown through RefMD's API.
Features
Streamable HTTP/SSE transport compatible with hosted MCP clients (e.g. Claude, Cursor)
Resource template
refmd://document/{id}to read document Markdown and metadataTools for listing, searching, creating, and updating documents via the RefMD API
Chunked document reads through optional
offset/limitparameters (each call returns up to 120,000 characters)Partial Markdown patching via
refmd-patch-document-content(insert/delete/replace without reuploading the entire file)
Prerequisites
Node.js 18+
Access to an existing RefMD instance
Either a personal access token (JWT) or credentials that can log in via
/api/auth/login
Configuration
The server now authenticates users via OAuth 2.1 with PKCE. Configure it with the following variables:
Variable | Description |
| Required. Base URL of your RefMD API (e.g.
). |
| Comma-separated list of allowed
values. Leave empty to allow any. |
| Comma-separated list of allowed redirect URIs. Defaults to allowing HTTPS URLs and
. Include ChatGPT’s callback URL (
). |
| Optional public issuer URL (defaults to the current request origin). |
| Optional access-token lifetime (default
). |
| Optional refresh-token lifetime (default
, i.e. 30 days). |
| Optional. Set to
,
, or
to persist OAuth tokens. Defaults to in-memory storage. |
| Optional connection string for the configured driver (e.g.
). For SQLite you may omit this and use
. |
| Optional filesystem path for SQLite storage (defaults to
). Accepts plain paths or
URLs; ensure the path resolves to persistent storage when using SQLite. |
/
| Optional listen port / host (defaults:
/
). |
Allowing multiple hosted clients: set
OAUTH_CLIENT_IDSto a comma-separated list (e.g.chatgpt-connector,Claude) and mirror the same set inOAUTH_ALLOWED_REDIRECTS. Leaving either variable empty keeps it open to any HTTPS redirect, but as soon as you specify one value you must list every connector you want to permit.
Remote MCP clients (Claude Web included) expect the OAuth Protected Resource Metadata document at
https://<host>/.well-known/oauth-protected-resourceso they can follow theresource_metadatahint inWWW-Authenticatechallenges. The server serves that document automatically (including mirrored aliases like/mcp/.well-known/...), so make sure your reverse proxy forwards those paths.
If you terminate TLS in a reverse proxy, make sure it forwards either the standard
Forwardedheader orX-Forwarded-Proto/X-Forwarded-Hostso the OAuth metadata advertises the correcthttps://origin. SetOAUTH_ISSUER=https://your-domainif you prefer an explicit override.
ℹ️ Install the appropriate database driver when enabling persistence:
npm install better-sqlite3for SQLite,npm install pgfor PostgreSQL, ornpm install mysql2for MySQL/MariaDB.
Install & Build
Run
The server exposes two transports:
http://<host>:<port>/sse— SSE transport (compatible with Claude SSE etc.)http://<host>:<port>/mcp— Streamable HTTP transport (one-shot POST per exchange)
OAuth flow
Configure your client (e.g. ChatGPT custom connector) with:
Authorization URL:
https://your-domain.example.com/oauth/authorizeToken URL:
https://your-domain.example.com/oauth/tokenRevocation URL:
https://your-domain.example.com/oauth/revokeScopes: (leave blank)
PKCE: enabled (ChatGPT uses S256 automatically)
When prompted, the browser shows the RefMD MCP consent page. Paste a RefMD API token generated from Profile → API tokens and approve.
The connector receives an access token and can call
/sseor/mcpwithAuthorization: Bearer <token>.
Tokens can be revoked from RefMD (profile page) or via POST /oauth/revoke.
Run with Docker
Mount a persistent volume (as shown above) so the SQLite database file survives container restarts.
Connecting a Chat Client
Claude (CLI):
claude mcp add --transport sse refmd https://your-domain.example.com/sseCursor / VS Code / MCP Inspector: choose an SSE transport and supply the same URL.
Once connected, resources appear under refmd://document/{id}. Available tools include refmd-list-documents, refmd-search-documents, refmd-create-document, refmd-read-document, refmd-update-document-content, and more.
Reading Large Documents
The refmd-read-document tool and the refmd://document/{id} resource both support optional pagination parameters so large Markdown files stay under the Model Context Protocol payload limits:
offset(default0): starting character position (zero-based).limit(default120000, capped at120000): maximum characters to return in the response.
Example resource URI: refmd://document/123e4567-e89b-12d3-a456-426614174000?offset=60000&limit=60000.
Each response includes range metadata and the next offset so clients can issue follow-up calls until the full document is retrieved.
Patching Document Content
Use the refmd-patch-document-content tool to insert, delete, or replace specific spans without sending the entire Markdown body:
Offsets/lengths are counted in Unicode code points to match RefMD’s editor behavior. The server validates ranges and applies the operations atomically before emitting document-change events, so downstream integrations stay in sync.
Release workflow
The GitHub Actions workflow CI MCP Server ships the container image. It runs automatically on pushes/PRs touching mcp-server and publishes to GHCR when:
the push is a tag matching
mcp-server-v*(versioned release), orthe workflow is manually triggered with
publish=true.
Tags published to ghcr.io/<owner>/refmd-mcp include semantic versions (1.2.0, 1.2, 1), the raw git tag, and latest. Use the extra-tag input for additional labels when invoking the workflow manually.
Development
Run in watch mode with TSX:
Any code changes require a rebuild (npm run build) before deploying or running with npm start.
This server cannot be installed