claude.ai Connectors

claude.ai's Connectors feature lets the web app talk to MCP servers running outside the browser. Papermark hosts an HTTP MCP endpoint so you don't need anything installed locally.

Setup

1. Open claude.ai and go to Settings → Connectors.
2. Click Add custom connector.
3. Set the URL to:

   https://mcp.papermark.com

4. Click Connect. claude.ai redirects you to Papermark to approve access (OAuth 2.1 device flow). Pick the scopes you want the agent to have, then approve.
5. claude.ai stores the OAuth token. You'll see papermark listed under your active connectors.

Verifying

Start a new conversation and check the Tools menu in the message input. papermark should be there with all 17 HTTP-transport tools (everything except upload_document, which requires local filesystem access and isn't exposed over HTTP).

Ask:

List my Papermark documents.

If auth went through, the agent calls list_documents and returns the result.

Re-authorizing

OAuth access tokens last 90 days. Connector refresh is automatic if you granted offline_access during setup (which is the default). If you ever explicitly revoked the connector or your refresh token expired, claude.ai will prompt you to reconnect — same flow as setup, just one click.

To revoke from your side:

• In claude.ai: Settings → Connectors → papermark → Disconnect
• In Papermark: Settings → Authorized Apps → claude.ai → Revoke

What's available

Same 18-tool surface as the stdio transport, minus upload_document (no filesystem access). See Tools for the full list.

Workflows that work especially well in the browser context:

• "Show me view analytics for the Q4 pitch deck."
• "Create a password-protected link to the investor memo and post the URL in this chat."
• "List every dataroom and its document count."
• "Who from Acme has viewed our materials this month?"

Limits

• No file upload. If you need to upload, use Claude Desktop or the CLI — both have local filesystem access.
• No state between sessions. Each conversation starts fresh. The connector persists; the chat context doesn't.
• Rate limits apply. Same per-token budget as the REST API (60 RPM by default). Heavy back-and-forth can hit it; the agent surfaces the rate_limit_exceeded error and waits.

Troubleshooting