Skip to main content
Harness Kit logoHarness Kit
DocsGitHub
Back to docs
workflowtavily
Install
Source: packages/harness-kit/src/registry/bundles/workflow/tavily/README.md

Tavily

Agent-oriented web access via Tavily MCP. The agent can search for current information, extract clean content from URLs, and escalate into site mapping or crawling when a single page is not enough.

What it installs

ArtifactPath (in your project)Purpose
MCP server.mcp.json -> tavily entryConnects the agent to Tavily MCP through a local npx server with your API key
Skill.agents/skills/tavily/Tool-selection protocol for search -> extract -> map -> crawl, query patterns, and noise-control guardrails

How it works

Tavily is more than a plain search endpoint. It is designed for agent workflows where the model needs to move from discovery to reading real pages without flooding context or burning time on broad crawls too early.

The MCP server gives the agent Tavily web access. The skill teaches when to use each capability:

  • tavily-search for current or domain-scoped discovery
  • tavily-extract when the exact URL is already known
  • tavily-map to inspect a site's structure before committing to deeper retrieval
  • tavily-crawl only when multiple related pages are clearly needed

This bundle currently installs Tavily as a local stdio MCP (npx tavily-mcp@latest) because that fits harness-kit's .mcp.json workflow. Tavily also offers a remote OAuth-capable MCP, but that would require a different artifact shape than the current installer supports.

Setup

  1. Get a Tavily API key from app.tavily.com
  2. Set TAVILY_API_KEY in your shell or project env
  3. Add the bundle: harness-kit add tavily

Optional: if you want default search behavior, set DEFAULT_PARAMETERS in your MCP env with a JSON object such as {"max_results":10,"search_depth":"advanced"}.

Pairs well with

  • crawl4ai - Tavily is strong for discovery and targeted extraction; crawl4ai is useful when you want raw site harvesting and larger scrape workflows
  • firecrawl - Tavily can find and narrow the right URLs first, then Firecrawl can handle heavier scrape or interaction flows
  • brave-search - alternate search provider if you want to compare result quality or use Brave's news/image/local features
  • quality-gates - when research drives code or docs changes, quality-gates ensures final claims are backed by fresh verification rather than search output alone