Plain semantic web search — the everyday search_web you reach for first. No category, no domain restriction: describe what you're looking for and Exa returns the closest pages. Calls the Exa REST API through a local script; no MCP server needed, only an EXA_API_KEY.
This skill is self-contained — its client (scripts/exa_client.py) is vendored, so you can install just this folder.
Install
Get a key first at https://dashboard.exa.ai/api-keys, then:
export EXA_API_KEY=your-key # add to ~/.zshrc / ~/.bashrc to make it permanent
Just this skill (Claude Code):
git clone --depth 1 https://github.com/CodeAlive-AI/exa-skills /tmp/exa-skills
cp -r /tmp/exa-skills/exa-web-search ~/.claude/skills/exa-web-search # user-level
# …or .claude/skills/exa-web-search inside a project for project-level
That's the whole install — no other folders required. (Other agents: drop the exa-web-search folder wherever your agent discovers skills.)
Whole collection: clone the repo and point your agent at it — see the repo README.
Run it
python scripts/web_search.py "<what you're looking for>" [-n N] [--text] [--json]
Examples:
python scripts/web_search.py "how do vector databases handle deletes" -n 10
python scripts/web_search.py "best practices for prompt caching with LLMs" --text
python scripts/web_search.py "self-hosted analytics alternatives to Google Analytics" -n 12
Output is a Markdown list (title · date · snippet) ending in sources_reviewed: N. --text includes page content, --json gives raw results, -h lists all flags.
Write queries that describe the page
Exa is semantic, not keyword. Describe the page you want to find, not the keywords. "detailed blog post about X written by a practitioner" beats "X". Run 2-3 different angles (not synonyms) in parallel for coverage.
Handy flags when you need them:
-c company|news|research paper|github|pdf|personal site|linkedin profile|financial report— focus a category without leaving this skill.--include-domains a.com,b.com/--start-published 2025-01-01— scope by site or date.--text— pull page content when snippets aren't enough.
More than search (same vendored client)
python scripts/exa_client.py contents <url> --text # read a page's full content
python scripts/exa_client.py answer "<question>" # a direct, cited answer
python scripts/exa_client.py similar <url> -n 10 # pages similar to a URL
Token isolation (for multi-search tasks)
For anything beyond 1-2 queries, dispatch a subagent per angle (use the Agent tool), have each run this script and return a compact list, then merge + deduplicate. Keeps raw search output out of your main context.
After you get results
- Treat results as similarity, not validation — skim and drop off-target hits before relying on them.
- Lead your answer with the finding; cite source URLs inline; prefer a table when fields are uniform.
Part of the Exa Skills collection. For specialized research (companies, people, papers, code, finance, personal sites, X) install the matching skill from the repo.