RSS MCP server

The tool surface

29 tools, each with a title and all four annotation hints so clients can render and gate them correctly. Your assistant picks the right one — you ask in plain language.

• Reading — get_unread_articles, get_new_articles, get_articles_by_feed, get_article, get_unread_count, get_article_counts
• Search — search_articles_by_keyword, search_articles_by_meaning (Max), get_story_clusters (semantic mode: Max)
• Watches — create_watch, list_watches, update_watch, delete_watch, get_watch_matches
• Feeds — get_feeds, get_feed, subscribe_to_feed, unsubscribe_from_feed, move_feed_to_folder, get_feed_health
• Curation — set_read_state, set_starred, get_starred_articles, get_reading_history
• Import / export — import_opml, export_opml, get_import_status
• Account — get_account_info, report_issue

get_new_articles is the one worth knowing about: it keeps a resumable per-client cursor, so "what's new since I last asked?" works across separate conversations without replaying everything you've already seen.

Technical details

• Endpoint — https://newsmind.app/mcp, one universal URL for every user.
• Transport — streamable HTTP with Server-Sent Events.
• Auth — OAuth 2.1: authorization code with PKCE S256, opaque tokens. Access tokens are short-lived and renewed through the refresh_token grant.
• Registration — dynamic client registration per RFC 7591 at POST /oauth/register, so compliant clients register themselves with no allowlist step.
• Discovery — /.well-known/oauth-protected-resource and /.well-known/oauth-authorization-server. Unauthenticated calls get a 401 with a WWW-Authenticate challenge carrying resource_metadata.
• Token fallback — for clients whose OAuth support is incomplete, pass a personal access token as Authorization: Bearer nm_….
• REST adapter — the same server speaks REST at /api/v1/*, with an OpenAPI 3.1 spec at /api/v1/openapi.json.

Add the RSS MCP server

1. Start a free trial — create a Newsmind account in under 30 seconds. 14 days free.
2. Add your feeds — paste RSS/Atom URLs or import an OPML file from any existing reader (Feedly, Inoreader, NetNewsWire, Reeder).
3. Add the server URL — point your MCP client at https://newsmind.app/mcp. OAuth signs you in; nothing to self-host.
4. Call the tools — your client can now read, search, cluster, and track your feeds through Newsmind's MCP tools.

See per-client setup →

Client compatibility

Clients that speak remote MCP over HTTP with OAuth connect with the URL alone. Everything else takes a bearer token.

• Claude — web, desktop, iOS, and Android, via custom connectors. Requires a paid Claude plan. See RSS for Claude.
• ChatGPT — supported through the same server. See RSS for ChatGPT.
• Gemini — added as a Spark custom app, web and mobile. Requires Google AI Pro or Ultra. See RSS for Gemini.
• Cursor, Cline, Zed, Continue, Gemini CLI — support varies by version, mostly in how completely each one implements the OAuth handshake. The bearer-token path works everywhere.

Most clients take a JSON config along these lines — substitute whichever key yours uses (mcpServers, context_servers, and modelContextProtocolServers are the common ones):

{
  "mcpServers": {
    "newsmind": {
      "url": "https://newsmind.app/mcp",
      "headers": {
        "Authorization": "Bearer nm_YOUR_TOKEN_HERE"
      }
    }
  }
}

If tools don't appear after configuring, the usual cause is that the client doesn't yet handle streamable-HTTP MCP or its OAuth flow is incomplete — try the bearer token before switching clients.

Why a hosted MCP server beats a DIY script

You could wire up your own RSS fetcher, but then you own the polling, the full-text extraction, dedupe, retention, auth, and uptime. Newsmind runs all of that as a hosted service: feeds are polled continuously, full article bodies are extracted (not just summaries), the same story across outlets is clustered, and a resumable cursor tracks what each client has already seen. Open standards throughout — MCP for native clients, OpenAPI for ChatGPT — so you're never locked to one assistant or one model.

Frequently asked questions