---
name: web-search-exa
description: Web search using Exa. Retrieves a clean context string and summarizes with sources.
triggers:
- web search
- browse web
- find sources
- research online
- summarize articles
- fact check
- latest updates
requires_mcp: exa
context: fork
---
# Web Search (Exa)
## Tool Restriction (Critical)
ONLY use `web_search_exa`. Do NOT use other Exa tools.
## Token Isolation (Critical)
Use Task agents for multi-query research; run single queries directly for quick lookups.
For multi-query research, spawn Task agents:
- Agent calls `web_search_exa`
- Agent deduplicates and extracts URLs + key facts
- Agent returns only distilled output (brief markdown or compact JSON)
- Main context stays clean regardless of search volume
## Inputs (Supported)
`web_search_exa` supports:
- `query` (string, required)
- `numResults` (number, optional; default: 8)
- `livecrawl` ("fallback" | "preferred", optional; default: "fallback")
- `type` ("auto" | "fast" | "deep", optional; default: "auto")
- `contextMaxCharacters` (number, optional; default: 10000)
## Dynamic Tuning
No hardcoded result counts. Tune dynamically:
- quick answer → numResults 3–5, type "fast"
- normal research → numResults 8–12, type "auto"
- comprehensive → numResults 20+, type "deep"
Live crawl:
- default: "fallback"
- if user asks "latest/current" → set "preferred"
## Query Variation
Generate 2–3 query variations for coverage, run in parallel, then merge + dedupe by URL.
## Output Format (Recommended)
Return:
1) Answer (short, structured)
2) Sources (URLs, 1-line relevance each)
3) "What's uncertain / conflicting" if needed
## MCP Configuration
```json
{
"servers": {
"exa": {
"type": "http",
"url": "https://mcp.exa.ai/mcp?tools=web_search_exa"
}
}
}
```
