Fetch recent arXiv papers in a category, sorted by submission date (newest first). days filters by published date.

Common categories: cs.AI (general AI), cs.LG (machine learning), cs.CV (computer vision), cs.CL (NLP), cs.HC (HCI / UX), cs.RO (robotics), cs.NE (neural networks), stat.ML (statistical ML), eess.IV (image/video processing — medical imaging lives here), eess.SP (signal processing), q-bio.QM (quantitative biology).

Search arXiv. Plain keywords work (auto-prefixed all:); for advanced queries use arXiv field syntax: ti: (title), au: (author), abs: (abstract), cat: (category, e.g. cat:eess.IV). days cuts off results older than N days (published field). When days is set, results are sorted by submission date instead of relevance.

Search PubMed for biomedical publications. Plain keywords work; for advanced queries use MeSH and field tags: mammography[MeSH], smith[Author], 2025[PDat]. Combine with AND/OR. days filters by publication date (PDat field).

Daily curated AI papers feed (now backed by Hugging Face's daily_papers — Papers with Code API was sunset after the 2024 HF acquisition). Empty query returns the newest curated papers. Search is client-side filtering over the daily-papers stream.

Browse github.com/trending — the public 'what's hot now' feed. USE THIS WHEN: user wants to browse trending repos with no specific topic in mind ('파이썬 트렌딩 보여줘', 'GitHub 핫한 거'). USE github_search INSTEAD WHEN: user has a specific topic/keyword. Note: this is HTML scraping (no official API), so layout changes can break it.

Search GitHub repositories by keyword via the official Search API. USE THIS WHEN: user has a specific topic ('medical imaging 리포', 'mammography GitHub'). days filters by repository created_at (treats it as 'repos created in the last N days') — pair with sort=stars for a stable trending-substitute. USE github_trending INSTEAD WHEN: no specific topic, just browsing.

Browse Hugging Face Hub. kind selects models / datasets / spaces (default models). sort: trending / downloads / likes / recent. days filters by lastModified — CAUTION: this catches old entries with recent edits, not just newly published ones. For 'truly new' discovery prefer sort='recent' + days=N.

Recent FDA 510(k) clearances via openFDA. Date filter is always applied. openFDA uses token-exact matching on string fields — for partial name matches use wildcards (e.g. device_name:mammo* not device_name:mammography).

Recent FDA medical device recalls via openFDA. Optionally filter by class (1=most serious, 3=least). Note: openFDA query syntax uses token-exact matching on string fields — for partial matches use wildcards (e.g. product_description:mammog*).

One-shot multi-source digest for a topic. Calls sources in parallel; partial failures don't break the report.

PRESENTATION RULES — follow strictly:

1. PRESERVE SECTION STRUCTURE. The output has separate per-source sections (📌 arXiv, 📌 PubMed, 📌 GitHub, etc.). Do NOT merge them.

2. TRANSLATE INLINE TEXT into the user's conversation language; keep section headers and emoji as-is.

3. PRESERVE VERBATIM: proper nouns, IDs (PMID, k_number, arXiv IDs), URLs, repository names, metric values.

4. Render every item — no summarization at the digest level.

Newspaper-style weekly briefing across all enabled sources. Topic is optional — without it, each source shows its 'what's new' feed. Use this when the user asks for '주간 뉴스' / '주간 트렌드' / 'weekly news' / 'briefing' style output.

PRESENTATION RULES — follow strictly:

1. PRESERVE STRUCTURE EXACTLY. The output is already organized into three groups (🎓 연구 동향 / 💻 코드 / 모델 / 🏥 규제 / 의료기기) and seven distinct source sections (arXiv, PubMed, Papers with Code, GitHub, Hugging Face, FDA 510(k), FDA Recalls). Do NOT merge sections (e.g. don't combine arXiv + PubMed into one 'papers' list). Do NOT reorder sections or items within a section. Do NOT change emoji or section headers.

2. TRANSLATE INLINE TEXT ONLY. Translate paper titles, abstracts, descriptions, and recall reasons into the user's current conversation language. Keep section headers, group titles, emoji, and metadata labels in their original form.

3. PRESERVE VERBATIM: proper nouns, author names, journal names, repository names (e.g. 'mattpocock/skills'), arXiv IDs, PMIDs, k_numbers, URLs, dates, and metric values (stars, downloads, etc.).

4. NO SUMMARIZATION at the briefing level. Render every item the tool returned. The user wants the full feed, not your synthesis.

5. ITEM-LEVEL DEPTH. For each paper, repo, model, or recall, preserve enough of the upstream abstract/description to convey what's new and why it matters — typically 2–4 sentences (around 150–300 chars of translated content per item). Do NOT collapse to a single headline-length sentence; the user wants to grasp each item without clicking through. Carry the problem → method → result/contribution structure when present in the source abstract.

Show current trends-mcp configuration: enabled sources, arXiv categories, PubMed default query, and which optional rate-limit tokens are set. Token VALUES are never returned (only whether they are configured). Use this to confirm state before or after a trends_set_* call.

Set which sources are enabled. Pass a list like ['arxiv', 'github']. Valid: arxiv, github, huggingface, paperswithcode, pubmed, fda_510k, fda_recalls. Pass ['*'] or ['all'] to enable all. Disabled sources' tools won't appear in the tool list at all (requires restart to take effect).

Set the default arXiv categories used by trends_briefing. Pass a list of entries; each entry is 'code' (e.g. 'cs.HC') or 'code:weight' (e.g. 'cs.HC:5'). Weight = papers per briefing per category (default 3). Example: ['cs.LG:5', 'cs.CV:3', 'cs.CL:2']. See ARXIV_CATEGORIES.md for the full list of valid codes.

Set the default PubMed query used by trends_briefing when no topic is provided. Use PubMed syntax: MeSH terms, [Title/Abstract] tags, AND/OR/NOT. [Title/Abstract] tags keep matches precise. Example: '(deep learning) AND (radiology[Title/Abstract])'.

Set or clear an optional rate-limit booster token. trends-mcp ONLY needs read access — when creating these tokens use the MINIMAL scope:

• github: NO scope at all (just authentication for rate limit). Do NOT use a token with 'repo' scope here.

• hf: read access only.

• ncbi / openfda: API keys (no scope concept). Pass empty string for value to remove a token. Provider must be one of: github, hf, ncbi, openfda.