Exa (formerly Metaphor)
Neural web search API purpose-built for LLM agents. Exa uses embeddings to search the web semantically — understanding concepts and relationships, not just keyword matching. Returns full page content (not just snippets) in clean Markdown format, enabling agents to actually read and process search results. Also provides FindSimilar (find URLs similar to a given URL), GetContents (fetch clean content from any URL), and neural search for specific content types (news, papers, tweets, GitHub).
Score Breakdown
⚙ Agent Friendliness
🔒 Security
HTTPS enforced. API key with no scope granularity. Startup company — SOC2 status not confirmed. No PII in standard search queries. Web content retrieval may surface sensitive cached content — consider privacy implications.
⚡ Reliability
Best When
LLM agents that need to search and retrieve web content with semantic understanding and clean Markdown output — eliminating the scraping/parsing pipeline between search and LLM consumption.
Avoid When
You need exact keyword matching (brand monitoring, patent search), very high volume at lowest cost, or real-time streaming news — Exa is optimized for semantic quality, not raw throughput.
Use Cases
- • Enable agents to search the web with semantic understanding — 'find papers on LLM alignment from 2024' returns conceptually relevant results, not just keyword matches
- • Retrieve full article content in clean Markdown via Exa's GetContents endpoint — ready for LLM context without HTML parsing
- • Find similar content to a known-good source using FindSimilar — 'find pages like this research paper'
- • Build RAG systems with fresh web content — Exa searches live web with LLM-optimized output, replacing SerpAPI + custom parsing
- • Search specific content types (news, papers, Twitter/X, GitHub) with type-specific search endpoints
Not For
- • Traditional keyword-based search where exact phrase matching matters — Exa's neural search is optimized for concept matching, not exact strings
- • High-volume web search at scale — Exa's pricing scales with requests; Bing API or Google SERP may be more cost-effective for high volume
- • Real-time news monitoring — Exa indexes web content but may have indexing lag; use dedicated news APIs for real-time monitoring
Interface
Authentication
API key in X-API-Key header. Key generated in Exa dashboard. Single key grants access to all API features. No scope granularity.
Pricing
Cost-effective for agent use cases — $0.01/search plus $0.001 per page content retrieved. GetContents is very cheap for bulk content retrieval. Credit card needed to add credits beyond free tier.
Agent Metadata
Known Gotchas
- ⚠ Exa's neural search is optimized for concepts, not keywords — agents must write queries as statements or concepts, not keyword lists ('papers about X' not 'X research paper keyword')
- ⚠ GetContents returns Markdown but quality varies by site — some sites return poor Markdown due to complex HTML structure; validate before sending to LLM
- ⚠ Search result freshness depends on Exa's crawl index — newly published content may not appear immediately
- ⚠ Content type filtering (type='news', type='papers') significantly improves result quality for specific domains — use type filters when applicable
- ⚠ numResults default is 10 — agents may need to increase for broader coverage, but more results mean more cost
- ⚠ MCP server enables direct Claude integration — agents using Claude can use the Exa MCP server for seamless tool integration
Alternatives
Full Evaluation Report
Detailed scoring breakdown, competitive positioning, security analysis, and improvement recommendations for Exa (formerly Metaphor).
Scores are editorial opinions as of 2026-03-06.