Exa Search CLI

Overview

Use exa search to discover and rank web sources. The command emits one JSON envelope to stdout and exits; successful API response fields are under data.

Set the credential in the environment:

export EXA_API_KEY="your_api_key_here"

Never put the credential in an argument, request object, source file, or log.

Quick Start

Default search:

exa search "latest developments in LLM evaluation"

The convenience form defaults to type: "auto", numResults: 10, and contents.highlights: true.

Bounded low-latency search:

exa search "recent AI infrastructure funding" \
  --type fast \
  --num-results 5 \
  --category news \
  --timeout-ms 30000

Use --pretty only for inspection. Automation should accept compact or pretty JSON rather than depend on whitespace.

Convenience Options

Choose:

• auto for general retrieval.
• fast for low-latency agent and product loops.
• instant for the lowest-latency interactive path.
• deep for multi-step discovery and synthesis.
• deep-reasoning for the most ambiguous research tasks.

Advanced Search Input

Use --input <file|-> when the request needs fields beyond the convenience options. - reads one complete Search request object from stdin:

exa search --input - <<'JSON'
{
  "query": "AI regulation policy updates",
  "type": "auto",
  "category": "news",
  "numResults": 10,
  "includeDomains": ["reuters.com", "bbc.com"],
  "startPublishedDate": "2026-01-01",
  "contents": {
    "text": {
      "maxCharacters": 2000
    },
    "maxAgeHours": 24,
    "livecrawlTimeout": 12000
  }
}
JSON

Do not mix --input with a positional query, --type, --num-results, or --category. Global flags such as --timeout-ms, --verbose, and --pretty remain valid.

Unknown request fields are forwarded, which lets new API fields work without a CLI release. The CLI validates the input as a JSON object; the upstream service remains authoritative for field combinations.

Search streaming is not exposed by this CLI. Omit stream; the Search command expects one JSON response.

Advanced Fields and Results

Read Search input when using result content, freshness, structured output, domain/date filters, subpages, or extras.

Read Search output before implementing envelope parsing, retry handling, costs, or grounding.

Critical Mistakes

• Do not place text, highlights, or summary at the request root; nest them in contents.
• Do not read results at the envelope root; read data.results.
• Do not use --input for a partial patch; it is the complete request object.
• Do not mix --input with convenience fields.
• Do not add Search streaming.
• Do not request multiple content modes without a concrete need.
• Do not assume every category supports every filter; let a rejected combination fail visibly and adjust the request.