exa-people-search

People Research with Exa

This skill uses Exa's neural search API to find people across LinkedIn profiles, personal sites, and the web. Exa's category: "people" mode returns structured profile data (name, title, company, location, experience) that you can filter and organize for the user.

Setup

The Exa MCP server must be configured. If it's not available, tell the user to run:

claude mcp add --transport http exa "https://mcp.exa.ai/mcp?tools=web_search_advanced_exa"

Then restart Claude Code.

Tool Restriction

ONLY use web_search_advanced_exa. Do NOT use web_search_exa or any other Exa tools - they lack the category and filtering parameters needed for people search.

Token Isolation

Never run Exa searches in main context. Always spawn Task agents to keep the main context clean:

• Agent runs the Exa search internally
• Agent processes results using LLM intelligence
• Agent returns only distilled output (compact JSON or brief markdown)

This matters because people searches can return large payloads with full profile text, and running multiple queries in parallel compounds this.

How to Search

Query Construction

Exa uses neural matching - describe what you're looking for in natural language rather than using keyword-style queries.

• Use singular form. "software engineer" not "software engineers"
• Describe what people do rather than exact titles. "engineer building voice AI" works better than "Voice AI Engineer"
• Combine role + company + location in one query. Example: "VP of Product at Stripe in London"
• Use OR for multiple roles. "CTO or VP Engineering or Head of AI at Databricks"

Dynamic Result Count (numResults)

Tune to user intent rather than hardcoding:

User says	numResults
"a few"	10-20
specific number	match it
1 company, 1 city	10-15
1 company, 2-3 cities	25
1 company, 5+ cities	50
multiple companies	25-50
"comprehensive" or "all"	50-100
ambiguous	ask the user

Higher counts return more results but also more noise. Since you filter for the user, erring on the high side is fine.

Query Variation for Coverage

Exa returns different results for different phrasings. For broad searches:

• Generate 2-3 query variations
• Run in parallel via separate agents
• Merge and deduplicate results

Categories

Use the appropriate Exa category depending on what you need:

Category	Use for	Notes
people	LinkedIn profiles, public bios	Primary for discovery. Limited filter support (see below)
personal site	Personal blogs, portfolios, about pages	Good for deeper individual research
news	Press mentions, interviews, speaker bios	Supports date filters
No category (type: "auto")	General web results, broader context	All filters available

Start with category: "people" for profile discovery, then use other categories for deeper research on specific individuals.

Filter Restrictions by Category

When using category: "people", these parameters cause errors - do not use them:

• startPublishedDate / endPublishedDate
• startCrawlDate / endCrawlDate
• includeText / excludeText
• excludeDomains
• includeDomains - only LinkedIn domains work (e.g., "linkedin.com"), and Exa already searches LinkedIn by default so this is rarely needed

When searching without a category, all parameters are available (but includeText/excludeText only support single-item arrays).

Common Pitfalls

1. Same-name companies. "Cohere" matches Cohere AI, Cohere Health, Cohere Consultants, and Cohere Technologies. "Anthropic" matches the AI lab and "Anthropic Softwares Pvt Ltd" in India. Always filter results to the correct entity.

2. Community vs. employees. Some companies have community programs (e.g., "Cohere Labs" researchers, "Claude Code Campus Ambassador"). These are community members, not employees. Filter accordingly when the user wants actual employees.

3. Location is not a hard filter. Exa uses neural matching, not exact filtering. A search for "engineer at OpenAI in London" may return OpenAI engineers in other cities too. Filter these out in post-processing - this is why higher numResults helps.

4. Profile truncation. At high result counts, individual profiles may be truncated. Name, title, and company are reliable; detailed work history may be cut off.

Examples

Discovery: find people by role

web_search_advanced_exa {
  "query": "VP Engineering AI infrastructure",
  "category": "people",
  "numResults": 20,
  "type": "auto"
}

With query variations for broader coverage

web_search_advanced_exa {
  "query": "machine learning engineer San Francisco",
  "category": "people",
  "additionalQueries": ["ML engineer SF", "AI engineer Bay Area"],
  "numResults": 25,
  "type": "deep"
}

Deep dive on a specific person

web_search_advanced_exa {
  "query": "Dario Amodei Anthropic CEO background",
  "type": "auto",
  "numResults": 15
}

News mentions with date filter

web_search_advanced_exa {
  "query": "Dario Amodei interview",
  "category": "news",
  "numResults": 10,
  "startPublishedDate": "2024-01-01"
}

GTM research: executive buyers at target accounts

web_search_advanced_exa {
  "query": "CTO or VP Engineering or VP Product at Stripe globally",
  "category": "people",
  "numResults": 50,
  "type": "auto"
}

Output Format

Return results as:

1. Results - name, title, company, location (when available)
2. Sources - profile URLs
3. Notes - profile completeness, verification status, any filtering applied

When the user asks for CSV or structured output, organize columns as: name, title, company, location, LinkedIn URL, and any additional fields they request.

Browser Fallback

Auto-fallback to Claude in Chrome browser when:

• Exa returns insufficient results
• Content is auth-gated (e.g., full LinkedIn profiles requiring login)
• Dynamic pages need JavaScript rendering