MCP Feed Reader

mcp-feed-reader-crunchtools

Secure MCP server for RSS/Atom feed reading with SQLite backend.

Installation

uvx (recommended)

uvx mcp-feed-reader-crunchtools

pip

pip install mcp-feed-reader-crunchtools

Container

podman run -v feedreader-data:/data quay.io/crunchtools/mcp-feed-reader

Configuration

Tools (17)

Feed Management

• add_feed_tool — Add an RSS/Atom feed by URL
• list_feeds_tool — List all feeds with unread counts
• get_feed_tool — Get feed details
• delete_feed_tool — Remove a feed
• refresh_feeds_tool — Crawl feed sources for new content (slow, prefer systemd timer)

Entry Management

• list_entries_tool — List entries (filterable, paginated)
• read_entry_tool — Read full entry content (auto-marks read)
• mark_read_tool — Mark entries as read
• mark_unread_tool — Mark entry as unread
• search_entries_tool — Full-text search (FTS5)

Category Management

• list_categories_tool — List categories with counts
• create_category_tool — Create category
• rename_category_tool — Rename category
• delete_category_tool — Delete category

Import/Export

• import_opml_tool — Import from OPML file
• export_opml_tool — Export as OPML
• get_stats_tool — Dashboard stats

Background Fetching

Nothing in the server crawls on its own. list_entries_tool and friends read a cache; if no one calls --fetch, that cache goes stale and every reader sees an empty feed with no error anywhere. A deployment without a crawl trigger is a broken deployment — this bit the daily briefing, which reported "no news" for days while all 61 feeds sat unfetched.

# Fetch all feeds via CLI
mcp-feed-reader-crunchtools --fetch

systemd timer (recommended)

deploy/ ships the two units used in production. Copy them to /etc/systemd/system/, adjust the container name in the service if yours differs, then:

sudo systemctl daemon-reload
sudo systemctl enable --now mcp-feeds-refresh.timer

The service runs --fetch inside the running container (podman exec), so the database keeps the SELinux category of the process that owns it. Do not bind-mount the data directory into a second container with :Z to run the crawl — that relabels the volume and locks the live server out.

The timer fires hourly, plus once at 05:45. That second entry is not redundant: consumers window entries on publication time, so an entry published between the last crawl and a 06:00 report is missing from that report and already older than the next day's window start — it is lost for good. Any consumer on a fixed schedule wants a crawl shortly before it.

Feeds are fetched conditionally (ETag / If-Modified-Since). A rate limit, gateway 5xx, or dropped connection is retried with exponential backoff, up to three attempts in all, before the feed is recorded as failed.

License

AGPL-3.0-or-later