Skip to content
mcp/skillhub

Compare · Web Search

Exa vs Tavily MCP Server

Exa exposes 2 tools and adds about 564 tokens to every model turn. Tavily exposes 4 tools and adds about 1,442. Tavily costs 2.6× as much context. Measured by connecting to each server and counting its tool list, not from the README.

Exaexa-labsTavilytavily-ai
Context tax~564 tok/turn(leaner)~1,442 tok/turn
Tools24
Cost per 1,000 turns (Claude Sonnet 5.5, uncached)$1.13$2.88
Security score97/10097/100
Cold start—241 ms
RunsLocal or hostedLocal (stdio)
Needs keysNone declaredTAVILY_API_KEY
FreshnessActiveActive
GitHub stars5.1k2.4k
LicenseMITMIT
Official MCP registryYesYes

Install in Claude Code

Exa

claude mcp add exa -- npx -y exa-mcp-server
Cursor, VS Code, Claude Desktop, Windsurf

Tavily

claude mcp add tavily -e TAVILY_API_KEY='<tavily-api-key>' -- npx -y tavily-mcp
Cursor, VS Code, Claude Desktop, Windsurf

Tools

By name, from each server's live tool list. The same name can still behave differently.

Only in Exa · 2

  • web_search_exa
  • web_fetch_exa

In both · 0

  • none

Only in Tavily · 4

  • tavily_search
  • tavily_extract
  • tavily_crawl
  • tavily_map

Questions

Which uses fewer tokens, Exa or Tavily?
Exa: its tool definitions take about 564 tokens per turn against 1,442 for Tavily (2.6× more). Tool definitions are resent on every model turn.
Do Exa and Tavily have the same tools?
No tool names overlap: Exa has 2 tools and Tavily has 4.
Are Exa and Tavily hosted or local?
Exa: local or hosted. Tavily: local (stdio).

Token counts use the o200k tokenizer on each server's tool definitions as an agent receives them; clients that load tools lazily pay less. Prices are Claude Sonnet 5.5 input at $2/MTok. Want to check one yourself? Run npx mcpskillhub test <package> — how it works.