research-feed

name: research-feed description: Manage and query the NightOwl research-feed plugin through its MCP tools — list/add/remove/enable sources (arxiv, reddit, bluesky, mastodon, googleSearch, googleScholar, chromeTabs), import open Chrome tabs (macOS), search adapters one-shot, browse cached items by relevance, save to citations, dismiss noise, prune. Use when the user wants to capture, organize, surface, or triage research material from these sources, or asks about "the feed" / "what I'm tracking" / "open tabs" / "recent papers" / "save this to citations".

This skill drives the research-feed MCP server (configured in .mcp.json). It exposes a small SQLite store of items pulled from research sources, plus an on-demand Chrome-tabs importer. Use the MCP tools directly; this skill encodes the workflow patterns and gotchas that aren't obvious from the tool descriptions alone.

When to invoke vs other approaches

If the user wants the rich panel UI (drag, click, AI relevance gating against current draft), tell them to open the IDE — npm run dev, then Cmd+, → 🧩 Plugins → toggle nightowl-research-feed → click the floating ⚲ button. The MCP tools cover the same data but without the AI gate.

Storage

All three surfaces (IDE plugin, CLI, MCP server) share one SQLite DB at ~/Library/Application Support/NightOwl/research-feed/feed.db on macOS — Electron's app.getPath('userData') for this app. Anything added through any surface shows up in the others.

Override per surface with RF_USER_DATA (env, set in .mcp.json for MCP; same env var or --user-data flag for the CLI). The IDE always uses Electron's path and is not configurable.

Recipes

Capture open Chrome tabs

feed_import_chrome_tabs (no args, or {appName: "Brave Browser"} etc.)

Returns {totalTabs, droppedByBlocklist, kept, newlyInserted}. Note this surface has no AI gate — it relies on the deterministic blocklist alone (chrome://, gmail, calendar, social DMs, banking, x.com/messages, etc.). The IDE's 📑 button does add an AI relevance gate against the current draft.

After import, offer to triage: feed_list_items({sourceId: "chrome-tabs", sort: "recent", limit: 20}) and walk the user through save/dismiss decisions.

First call may trigger a macOS Automation prompt for whichever app is controlling — usually the terminal Claude Code is running in. If denied, the tool returns a friendly error explaining how to grant it.

Find recent research without subscribing

For ad-hoc lookups where the user doesn't want a polling source:

feed_search({type: "arxiv", config: {category: "cs.LO", maxResults: 10}})
feed_search({type: "bluesky", config: {query: "phenomenology", lang: "en", limit: 20}})
feed_search({type: "reddit", config: {subreddits: ["philosophy"], sort: "new", limit: 15}})

Results come back inline; nothing is written. Show titles + authors + URLs; offer to subscribe via feed_add_source if anything looks like a recurring interest.

Start tracking a new topic

feed_add_source({
  id: "arxiv-cogsci",
  type: "arxiv",
  config: {category: "cs.AI", query: "embodied cognition", maxResults: 20}
})

Sensible defaults per type:

• arxiv: {category: "cs.AI" | "cs.LG" | "cs.LO" | ..., query?: string, maxResults: 20} — categories are arxiv subject codes
• reddit: {subreddits: [...], sort: "new", limit: 20} — User-Agent is set internally
• bluesky: {query: "...", lang: "en", limit: 20}
• mastodon: {instances: ["mastodon.social", "scholar.social"], tags: ["..."], limit: 20}
• googleSearch: {query: "...", cx: "<user's CX>", dateRestrict: "w1"} — costs ~$5/1000, requires API key in IDE
• googleScholar: {query: "..."} — paid SerpAPI, off by default

Default poll interval is 15 minutes. Pick descriptive IDs: arxiv-philosophy-of-mind not src1.

Triage the feed

feed_list_items({sort: "score", limit: 20, dismissed: false})

Sort by score if items have AI scores from the IDE; sort by recent otherwise (MCP-added items have no scores). Show a numbered list. For items the user is unsure about, fetch detail with feed_get_item({id: N}) before deciding. Batch save/dismiss in one go: e.g. user says "save 1, 3, 7, dismiss the rest" → fire the calls in parallel.

Inspect or debug a source

If a source shows errors:

• feed_list_sources — find lastError for the source
• feed_search with the same type and config — reproduce out-of-band; the error message will be more direct than what's stored
• feed_set_enabled({id, enabled: false}) to stop the poll loop from retrying while you investigate

For Chrome tabs returning unexpectedly few items, the deterministic blocklist may be hiding work tabs. Ask the user if they have research on, e.g., github.com/notifications, then suggest extraBlocklist: [] plus a custom set, or blocklist: [...] to replace entirely.

After-action behavior

• After feed_import_chrome_tabs or feed_fetch, proactively ask if the user wants to triage the new items. Don't just dump the count.
• After feed_save_item, mention that the IDE panel saves into the citations DB but the MCP path only marks saved=true in the feed store (citations integration is IDE-only). If the user expected a citation entry, they'll need to do that step from the panel.
• For destructive actions (feed_remove_source, feed_prune), confirm scope before firing. Removing a source deletes its cached items too.

Quick reference — the 12 tools

feed_list_sources, feed_add_source, feed_remove_source, feed_set_enabled
feed_fetch, feed_search
feed_list_items, feed_get_item, feed_save_item, feed_dismiss_item, feed_prune
feed_import_chrome_tabs

For exact schemas, the MCP server publishes them on tools/list — Claude Code already has them loaded, no need to re-fetch.