Web search and page fetching. Returns {success, results/content, url}. Use ONLY after exhausting local sources: retrieve() -> kg(neighbors) -> web. Actions: 'search' (DuckDuckGo, returns title/url/snippet), 'crawl' (fetch page content). Set news=true for DDG news endpoint (adds date/source fields). High context cost - prefer harvest=True to save content to disk instead of loading into context. Results cached 20min. For bot-protected sites, use fx tool instead.

Wikipedia lookup. Check memory/KG for known facts first. Actions: search, page. Results cached 1hr.

GitHub file and repo access via API. Returns {success, content/stats/results}. Use for: fetching source files, analyzing repo structure, searching code across GitHub. Actions: 'fetch' (raw file content), 'analyze' (repo stats/structure), 'search' (code search), 'create_issue' (open new issue), 'comment_issue' (add comment), 'get_issue' (read issue + comments), 'list_issues' (list repo issues), 'update_issue' (edit state/title/body/labels). Requires GITHUB_TOKEN env var for write ops and higher rate limits. 2MB response cap, 10min cache. For large files, use harvest=True to save to disk instead of loading into context.

YouTube transcript and metadata tool. Returns captions + rich metadata for any YouTube video. Actions: 'transcript' (captions only), 'metadata' (title/channel/duration/description/chapters only), 'info' (transcript + metadata in parallel - the common case), 'list_languages' (enumerate available caption tracks), 'search' (find a phrase within a transcript, returns timestamped matches with t= deeplinks). Primary fetch: youtube-transcript-api (fast, clean). Fallback: yt-dlp with json3 subtitles + VTT roll-up dedup. Format options: text (default, lowest tokens), timestamped, srt, vtt, json. Use harvest=True to write clean markdown under harvested/youtube/ for RAG ingestion. Accepts any URL shape (watch, youtu.be, shorts, live, embed) or bare 11-char video_id. Cached 1hr. Auto-paginates long transcripts; use segment_offset to continue from where the last response ended. For long podcasts, prefer harvest=True then retrieve() instead of loading full text into context.

Wikidata SPARQL queries. Use for structured entity data (people, places, organizations, factual relationships). 5000 char query limit, 15s timeout, 1000 max results. Consider storing results in KG for reuse.

Anti-detect Firefox browser. Use when web(crawl) fails due to bot protection (Cloudflare, Datadome, Reddit verification). Persistent profile at data/fx_profile/: cookies, localStorage, and solved Cloudflare challenges survive MCP reboots. Fast defaults: images blocked, humanize off, domcontentloaded wait. Prefer 'read' for one-shot extraction — auto-waits for JS challenges (networkidle), runs trafilatura, auto-closes. Supports harvest=True to save content to disk instead of loading into context. Use 'navigate' + explicit actions for multi-step workflows (logged-in flows, form fills, clicks); call 'close' when done. Actions: navigate, read, click, fill, type, screenshot, extract, text, wait, press, scroll, evaluate, cookies, add_cookies, close, status.