Skip to main content
Agent Skills add Tavily’s web capabilities to AI coding agents like Claude Code, Cursor, Cline, Codex, Windsurf, and others. They run the Tavily CLI, and the source is in tavily-ai/skills.
1

Install the Tavily CLI

See the CLI docs for other installation methods.
2

Install the skills

Or install a specific skill:
3

Restart your agent

Restart your AI agent to load the skills.

​
Available Skills

The core skill gives your agent Tavily’s tools. Each task skill uses the core skill to produce a specific result, such as a report or a table, and installs it if it’s missing.

​
Usage

Once installed, skills are automatically available to your agent. No additional configuration is needed — your agent will use them when appropriate based on your prompts.

​
Automatic Invocation

Simply describe what you need and your agent will use the right Tavily skill:

​
Explicit Skill Invocation

You can also invoke skills directly using slash commands:

​
Skill Details

​
Core skill

tavily-search covers search, extract, map, and crawl. Your agent picks the tool from your prompt, so all four are invoked with /tavily-search.

Search

Web search returning LLM-optimized results with content snippets and relevance scores.Invoke explicitly:
Example prompts:
  • “Search for the latest news on AI regulations”
  • “/tavily-search current React best practices”
  • “Search for Python async patterns”
CLI usage:
Key options: --depth (ultra-fast/fast/basic/advanced), --max-results, --topic, --time-range, --include-domains, --exclude-domains, --include-raw-content
Extract clean markdown or text content from one or more URLs. Handles JavaScript-rendered pages.Invoke explicitly:
Example prompts:
  • “Extract the content from this article URL”
  • “/tavily-search extract the main content of https://example.com/blog/post”
  • “Extract content from these three documentation pages”
CLI usage:
Key options: --query, --chunks-per-source, --extract-depth (basic/advanced), --format (markdown/text)
Discover URLs on a website without extracting content. Faster than crawling — useful for finding the right page before extracting.Invoke explicitly:
Example prompts:
  • “Map the site structure of docs.example.com”
  • “Find the authentication page on this site”
  • “List all URLs under the /api/ path”
CLI usage:
Key options: --max-depth, --limit, --instructions, --select-paths, --exclude-paths
Use the map + extract pattern: map to find the right page, then extract its content. This is often more efficient than crawling an entire site.
Crawl websites and extract content from multiple pages. Save each page as a local markdown file or get structured JSON output.Invoke explicitly:
Example prompts:
  • “Crawl the Stripe API docs and save them locally”
  • “/tavily-search crawl https://docs.example.com”
  • “Download the Next.js documentation for offline reference”
CLI usage:
Key options: --max-depth, --limit, --instructions, --chunks-per-source, --output-dir, --select-paths, --exclude-paths

​
Task skills

Task skills run the core skill’s tools in a fixed workflow and return a finished result. Use one when you want a report, a record, a table, or a recurring check rather than raw search results.
Researches a topic in depth. The agent splits the question into subtopics that don’t overlap, researches them in parallel, using subagents when the agent supports them, fills in any thin sections, and writes a cited report. Use tavily-search instead when you need a single fact.Invoke explicitly:
Example prompts:
  • “Research the competitive landscape for AI coding assistants”
  • “Compare the EU and US approaches to AI regulation across liability, transparency, and enforcement”
  • “/tavily-research AI agent frameworks and save the report to a file”
Output: a cited report in the reply. When you ask for a file, it’s saved to reports/{title}.md.
Fills in a CRM record for one company, person, or organization, or adds fields to a CSV of rows you already have. The agent first confirms which entity it is by matching two identifiers, such as the name and the company’s domain, so a record never mixes two entities that share a name. Every value comes from a page the agent read, and a field with no source stays empty.Invoke explicitly:
Example prompts:
  • “Enrich Ramp (ramp.com) for our CRM”
  • “Who is the CTO of Linear, and where did they work before?”
  • “Fill in industry, headquarters, and employee count for every company in accounts.csv”
Output: a record with a source for each field, plus key people and recent news. If more than one entity still fits, the agent returns the candidates instead of picking one.
Builds a deduplicated table of matches for a set. Before searching, the agent confirms the inclusion rule, the columns, and a cap. It then searches from several angles, such as the category, roundup pages, and directories, and merges rows that share a website.Invoke explicitly:
Example prompts:
  • “List Series A developer-tools startups in Europe, capped at 50”
  • “Find every open-source vector database, with license and GitHub stars”
  • “/tavily-list climate-tech accelerators in the US”
Output: the table in the reply, also saved to lists/{title}.md with the angles searched and the likely gaps. The table is what the searches returned, not every match that exists. Ask for more and the agent searches again and appends only new rows.
Sets up a search or an extraction that runs again on a cadence. A search monitor alerts on result URLs that weren’t in the previous run. An extract monitor alerts when a known page’s content changes, and quotes the part you care about, such as a price or a version number. Search monitors set max_age_hours to the cadence, so cached results older than the last run aren’t returned.Invoke explicitly:
Example prompts:
  • “Watch for news mentioning our company, daily”
  • “Tell me when the pricing page at https://example.com/pricing changes”
  • “Track new job postings for staff ML engineers at Anthropic, weekly”
Output: a monitors/{name}/ folder with monitor.md (the query or URLs, the cadence, and the change rule) and snapshot.json (the baseline from the first run). The skill doesn’t start a scheduler unless you ask. Point your scheduler, such as Claude Code’s /loop or cron, at the folder.

​
Developer skill

Build production-ready Tavily integrations with best practices baked in. Reference documentation for implementing web search, content extraction, crawling, and research in agentic workflows, RAG systems, or autonomous agents.Invoke explicitly:
Example prompts:
  • “Add Tavily search to my internal company chatbot so it can answer questions about our competitors”
  • “Build a lead enrichment tool that uses Tavily to find company information from their website”
  • “Create a news monitoring agent that tracks mentions of our brand using Tavily search”
  • “Implement a RAG pipeline that uses Tavily extract to pull content from industry reports”

​
What You Can Build

Copy-paste these prompts into your AI agent and start building:
Build a chatbot that can answer questions about current events and up-to-date information.Try these prompts:
Create a live news dashboard that tracks topics and analyzes sentiment.Try these prompts:
Build tools that automatically enrich leads with company data from the web.Try these prompts:
Build an autonomous agent that monitors competitors and surfaces insights.Try these prompts:
The /tavily-sdk skill is your fastest path to production. Describe what you want to build and your agent generates working code with best practices baked in.