Tavily

@Tavily search for recent benchmarks comparing Claude 3.5 Sonnet and GPT-4o on coding tasks, limit to 5 results from technical blogs

Tavily returned 5 articles published in the last 60 days. Top result: a SWE-bench analysis from Scale AI showing Claude 3.5 Sonnet resolving 49% of issues vs. GPT-4o's 38%. Second result: an Aider leaderboard update noting Claude's stronger refactoring accuracy. Third: a HumanEval comparison from Papers with Code. Fourth: a Reddit discussion thread aggregating community benchmarks. Fifth: a blog post from Anthropic detailing their internal eval methodology. All links verified and accessible.

This example uses domain filtering and result-count controls to narrow a broad query. Tavily's search depth defaults to 'basic' unless you specify 'advanced', which costs more tokens but returns richer context. Results reflect the index state at query time — fast-moving topics may show older content if the web hasn't updated.

@Tavily find the latest FDA approval announcements for GLP-1 drugs, search depth advanced, include only .gov domains

Tavily performed an advanced search restricted to .gov domains and returned 3 results. First: an FDA press release from two weeks ago announcing expanded approval for semaglutide in adolescents. Second: a clinical trial summary posted to ClinicalTrials.gov detailing a Phase 3 study for tirzepatide. Third: an FDA safety communication regarding supply constraints for dulaglutide. Each result includes a content snippet and publication date. No commercial or news sites appeared due to the domain filter.

Advanced search depth increases latency and token cost but surfaces more authoritative excerpts. Domain filtering to .gov ensures regulatory sources only — useful when you need primary documentation rather than news coverage. Be aware that FDA.gov updates lag behind press cycles by days.

@Tavily search for case studies on companies using MCP servers in production, then summarize the common implementation patterns and pain points mentioned

Tavily found 4 case studies and 2 technical blog posts. Common patterns: teams start with read-only MCPs (Slack, GitHub) before adding write capabilities; most run MCP servers in Docker containers with API key rotation; several mention using ngrok or Tailscale for local dev. Pain points: OAuth refresh token handling cited in 3 sources; rate-limit coordination across multiple AI agents in 2 sources; one case study noted difficulty debugging MCP transport errors without structured logging. The synthesis suggests starting with a single MCP and adding observability before scaling to multi-server setups.

This example pairs Tavily's search with the AI's summarization. The MCP retrieves raw content; the AI extracts themes. Result quality depends on how well the search query matches available content — 'MCP servers' is a niche topic, so fewer high-quality sources exist compared to mainstream queries. Tavily does not deduplicate overlapping articles.

When Tavily beats internal search for support teams

A 6-person support team fields 40 tickets a day about product integrations and compliance questions. Half the answers live in vendor docs or third-party sites, not the internal wiki. Tavily search runs in Switchy's shared workspace so the whole team sees the same results when someone asks "does Stripe support ACH debits in Canada" or "what's the GDPR retention rule for session logs." The search depth control matters here: basic mode for quick checks during live chats, advanced mode when writing help articles that need citations. The domain filter keeps results focused on official sources instead of Reddit threads. If your team already maintains a comprehensive internal knowledge base, you don't need this. But if you're constantly Googling vendor specs or regulatory details mid-ticket, Tavily in Switchy cuts 3-5 minutes per lookup and keeps the search history visible to the next person who needs it.

Tavily for ad-hoc competitor intel gathering

A 3-person product team runs monthly competitor reviews before roadmap planning. They need to pull recent feature announcements, pricing changes, and customer sentiment from across the web in 90 minutes. Tavily search in Switchy lets them queue 8-10 queries ("[competitor] new features 2025", "[competitor] enterprise pricing") and review results together in one workspace instead of Slacking screenshots back and forth. The result count parameter (default 5, max 20) means they can scan quickly or go deep depending on the topic. The content type filter (news, articles, forums) helps separate official announcements from user complaints. This works when you're researching 2-4 competitors in a focused session. If you're tracking 15 competitors daily, you need a monitoring tool, not a search MCP. For quarterly or monthly sprints where the team needs fresh web data fast, Tavily in Switchy beats manual Googling by 40 minutes per session.

When Tavily helps sales reps answer technical objections

A 5-person sales team closes deals that hinge on technical integration questions: "Does your product work with Salesforce CPQ?" or "What's your uptime compared to [competitor]?" The rep needs an answer in the next 15 minutes, and the internal wiki is three months out of date. Tavily search in Switchy's shared workspace means the rep can query "[our product] Salesforce CPQ integration 2025" and get current results from vendor docs, community forums, and third-party reviews. The search depth toggle (basic for quick checks, advanced for due diligence) matches the urgency. The domain filter keeps results on-brand when they need to cite sources in the follow-up email. This works for teams closing 10-20 deals a month where objections vary deal-to-deal. If your objections are repetitive, build a battle card. If they're one-off technical questions that require current web data, Tavily in Switchy gets the rep an answer before the prospect's next meeting.