Exa Search

High-precision semantic search via Exa API. Standalone CLI only (no MCP dependency).

Execution Method

# Prerequisites: pip install httpx tenacity
# Environment: EXA_API_KEY (required), EXA_API_URL (optional, default: https://api.exa.ai)

# All examples assume cwd == skills/exa/. The shim auto-chdirs if you launch
# it from elsewhere (e.g., the repo root).
cd skills/exa
python scripts/exa_cli.py --help

Available Tools

# Basic semantic search (highlights always on; supports inline category:<type>)
python scripts/exa_cli.py web_search_exa --query "TypeScript design patterns" [--num-results 10]
python scripts/exa_cli.py web_search_exa --query "category:company Anthropic AI safety"

# Batch URL fetch (urls is a repeatable flag; payload field is upstream `ids`)
python scripts/exa_cli.py web_fetch_exa \
  --urls "https://a.com" --urls "https://b.com" \
  [--max-chars 3000] [--out content.json]

# Advanced filtered search (list params are repeatable flags, no comma syntax)
python scripts/exa_cli.py web_search_advanced_exa --query "transformer" \
  [--type auto|fast|instant] [--category research\ paper] \
  [--include-domains arxiv.org --include-domains papers.nips.cc] \
  [--exclude-domains medium.com] \
  [--include-text "attention"] [--exclude-text "tutorial"] \
  [--start-date 2024-01-01] [--end-date 2024-12-31] \
  [--num-results 10] [--max-age-hours 168] \
  [--text] [--highlights] [--summary] \
  [--max-chars 5000]   # only effective when --text is set; emits stderr warning otherwise
  [--out results.json]

# Configuration / connectivity probe (omit --no-test to run a numResults=1 ping)
python scripts/exa_cli.py get_config_info [--no-test]

# Create an Agent run or resume the same retained run
python scripts/exa_cli.py agent_run --query "Research an evidence-backed market map" [--effort low]
python scripts/exa_cli.py agent_run --run-id agent_run_123 [--wait-seconds 750]

Tool Capability Matrix

Global Options

Place before the subcommand:

Tool Routing Guide

Workflow Patterns

Pattern 1: Quick Semantic Search

python scripts/exa_cli.py web_search_exa --query "best practices for React hooks" --num-results 5

Pattern 2: Filtered Research (repeatable flags)

python scripts/exa_cli.py web_search_advanced_exa --query "transformer architecture" \
  --include-domains arxiv.org --include-domains papers.nips.cc \
  --start-date 2023-01-01 --text --summary

Pattern 3: Batch URL Read

python scripts/exa_cli.py web_fetch_exa \
  --urls "https://example.com/a" --urls "https://example.com/b" \
  --max-chars 4000 --out batch.json

Pattern 4: Third-Party Endpoint (Bearer Auth)

# Connect to exa-pool or other Exa-compatible proxy
export EXA_API_URL=https://pool.example.com
export EXA_AUTH_SCHEME=bearer
export EXA_API_KEY=your-bearer-token

python scripts/exa_cli.py web_search_exa --query "AI agents" --num-results 5

# Or use CLI flags for one-off requests
python scripts/exa_cli.py --auth-scheme bearer --api-url https://pool.example.com \
  web_search_exa --query "AI agents"

Pattern 5: Exa Agent

Read [exa-agent.md](exa-agent.md) before creating a run. It defines objective/schema/coverage checks, --run-id versus --previous-run-id, evidence validation, and ZDR limits without duplicating that workflow here.

References

The references/ directory carries 11 prompt-engineering guides. Open them on demand when crafting queries:

Error Handling

Output Format

All commands print JSON (ensure_ascii=False, indent 2) to stdout. With --out <file>, the response JSON is written to that path and stdout becomes {"status":"ok","file":"<file>"}. Errors go to stderr as {"error":"<message>"} with non-zero exit.

Agent creation and interruption additionally emit one-line JSON progress events to stderr so the retained run ID remains recoverable without corrupting stdout.