Company Research

Company Research

Tool Selection (Critical)

Two Exa surfaces, two jobs:

• Exa Agent (agent_run) — the default for company research. Use it for deep dives, competitor analysis, multi-angle research (product + funding + news + people), and building company lists. One Agent run handles query decomposition, multi-step searching, and synthesis internally — do not orchestrate many manual searches for work an Agent run covers.
• web_search_advanced_exa — quick, low-latency lookups: a fast category: "company" discovery pass, a single news check, or finding a homepage.

Do NOT use other Exa tools.

Deep Dives and Lists: Exa Agent

Agent runs may stream to completion in one call. If a run outlives the MCP call window, continue waiting with its returned run ID.

1. Call agent_run with a natural-language query and, when you want repeatable structure, an outputSchema (bound arrays with maxItems).
2. If it returns status: "running" with a runId, call agent_run again with only that runId until outputReady is true.
3. Read output.text or output.structured, plus output.grounding citations, from the agent_run result.

Useful inputs: systemPrompt (source preferences, dedup rules), input.exclusion (companies to avoid), previousRunId (a new follow-up run based on a completed run), effort ("low" default; "auto" or "high" for more depth).

Example: company deep dive

Copy code
agent_run {
  "query": "Research Anthropic: product lines, funding history and valuation, key executives, main competitors, and notable news from the last 6 months.",
  "effort": "auto",
  "outputSchema": {
    "type": "object",
    "properties": {
      "overview": { "type": "string" },
      "funding": { "type": "array", "maxItems": 10, "items": { "type": "object", "properties": { "round": { "type": "string" }, "amount": { "type": "string" }, "date": { "type": "string" } }, "required": ["round"] } },
      "competitors": { "type": "array", "maxItems": 10, "items": { "type": "string" } },
      "key_people": { "type": "array", "maxItems": 10, "items": { "type": "object", "properties": { "name": { "type": "string" }, "title": { "type": "string" } }, "required": ["name", "title"] } }
    },
    "required": ["overview", "competitors"]
  }
}

Example: build a company list

Copy code
agent_run {
  "query": "Find 25 AI infrastructure startups headquartered in San Francisco. For each, include what they build and their latest funding stage.",
  "effort": "auto",
  "outputSchema": {
    "type": "object",
    "properties": {
      "companies": {
        "type": "array",
        "maxItems": 25,
        "items": {
          "type": "object",
          "properties": {
            "name": { "type": "string" },
            "website": { "type": "string", "format": "uri" },
            "description": { "type": "string", "description": "in 12 words or less" },
            "funding_stage": { "type": "string" }
          },
          "required": ["name", "website", "description"]
        }
      }
    },
    "required": ["companies"]
  }
}

Quick Lookups: Advanced Search

Use web_search_advanced_exa when a single fast search answers the question. Tune numResults to intent (a few → 10-20; comprehensive → 50-100; specified → match it).

Categories

• company → homepages, rich metadata (headcount, location, funding, revenue)
• news → press coverage, announcements
• people → public professional profiles
• No category (type: "auto") → general web results, broader context

Default to type: "auto". Prefer highlights for content extraction; do not stack text + highlights + summary in one call.

Category-Specific Filter Restrictions

Unsupported category/filter combinations return 400 errors:

• category: "company" does not support published-date or crawl-date filters, excludeDomains, or exact-text filters; express constraints like "founded after 2020" in the query instead
• category: "people" does not support published-date, crawl-date, domain, or exact-text filters; put all filtering in the natural-language query
• Without a category (or with news), domain and date filters work fine

Examples

Discovery pass:

Copy code
web_search_advanced_exa {
  "query": "AI infrastructure startups San Francisco",
  "category": "company",
  "numResults": 20,
  "type": "auto"
}

News check:

Copy code
web_search_advanced_exa {
  "query": "Anthropic AI safety",
  "category": "news",
  "numResults": 15,
  "startPublishedDate": "2025-01-01"
}

Key people:

Copy code
web_search_advanced_exa {
  "query": "VP Engineering AI infrastructure",
  "category": "people",
  "numResults": 20
}

Token Isolation

Never dump raw search results into main context. Spawn Task agents for Advanced Search calls; for Agent runs, go straight from output.structured to the final answer.

Browser Fallback

Fall back to Claude in Chrome only when content is auth-gated or requires JavaScript rendering.

Output Format

Return:

1. Results (structured list; one company per row)
2. Sources (URLs; 1-line relevance each — use output.grounding from Agent runs)
3. Notes (uncertainty/conflicts)

References

• Exa Agent guide: https://exa.ai/docs/agent/quickstart
• Company Search reference: https://exa.ai/docs/search/data/companies-people
• Exa MCP setup: https://exa.ai/docs/get-started/exa-mcp
• Full docs for LLMs: https://exa.ai/docs/llms.txt