exa

name

exa

description

Search the web, extract content from URLs, and get AI-powered answers using the Exa API via curl. Use this skill whenever you need real-time web information: researching a topic, finding documentation, looking up people or companies, reading news, fetching page content, or answering questions with citations. Also use when the user says /exa, "search the web", "look this up", "find articles about", "what's the latest on", or any query that requires up-to-date web knowledge beyond your training data.

Exa Web Search Skill

Search the web, extract page content, and get cited answers — all via curl against the Exa API.

Authentication

The API key is read from the environment variable EXA_API_KEY. Every curl call uses:

-H "x-api-key: $EXA_API_KEY"

If EXA_API_KEY is not set, stop and tell the user to set it:

export EXA_API_KEY="their-key-here"

How to Use This Skill

When you need web information, pick the right endpoint and search type for the task, build the curl command, execute it, then parse and present the results clearly to the user. Always attribute sources with URLs.

Endpoints

1. /search — Find and optionally retrieve content

The primary endpoint. Finds relevant pages and optionally returns their content.

curl -s -X POST 'https://api.exa.ai/search' \
  -H "x-api-key: $EXA_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
  "query": "your search query here",
  "type": "auto",
  "num_results": 5,
  "contents": {
    "highlights": { "max_characters": 4000 }
  }
}'

2. /contents — Extract content from known URLs

Use when you already have URLs and need their content. Content options (text or highlights) go at the top level (not nested in contents).

curl -s -X POST 'https://api.exa.ai/contents' \
  -H "x-api-key: $EXA_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
  "urls": ["https://example.com/article"],
  "highlights": { "max_characters": 4000 }
}'

3. /answer — Q&A with citations

Get a direct answer to a question, backed by web sources.

curl -s -X POST 'https://api.exa.ai/answer' \
  -H "x-api-key: $EXA_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
  "query": "What is the current status of Project X?",
  "text": true
}'

Search Types

Pick the type based on the user's need:

Type	When to use	Speed
auto	Default choice. Balanced relevance and speed	~2s
fast	Quick lookups, autocomplete-style queries	<1s
deep	Research, enrichment, thorough results. Supports outputSchema	4-12s
deep-reasoning	Complex multi-step reasoning across many sources. Supports outputSchema	10-30s

Default to auto unless the user needs depth or structured output.

Content Options

Choose ONE per /search request (nested inside "contents"):

Option	Config	When to use
Highlights	"highlights": {"max_characters": 4000}	Default. Summaries, Q&A, general research — token-efficient
Text	"text": {"max_characters": 10000}	Full articles, code, documentation — higher token cost

For /contents endpoint, these go at the top level (not inside contents).

Default to highlights to keep token usage low. Use text only when the user needs full content (code examples, complete documentation, etc.).

Structured Outputs (Deep Search Only)

When type is deep or deep-reasoning, you can request structured JSON via outputSchema. Exa returns web-grounded data with field-level citations.

curl -s -X POST 'https://api.exa.ai/search' \
  -H "x-api-key: $EXA_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
  "query": "AI startups in healthcare France",
  "type": "deep",
  "outputSchema": {
    "type": "object",
    "required": ["companies"],
    "properties": {
      "companies": {
        "type": "array",
        "description": "Companies found",
        "items": {
          "type": "object",
          "required": ["name"],
          "properties": {
            "name": { "type": "string" },
            "description": { "type": "string" }
          }
        }
      }
    }
  },
  "contents": { "highlights": { "max_characters": 4000 } }
}'

Response shape:

• output.content — structured JSON matching your schema
• output.grounding — field-level citations with URLs and confidence

Category Filters

Optional. Search dedicated indexes for specific content types:

Category	Query style	Notes
"people"	Role/expertise description, singular form	No date/text filters
"company"	Industry/attributes, singular form	Returns company objects
"news"	Topic/event	Use maxAgeHours: 24 for breaking news
"research paper"	Technical topic	Includes arxiv, paperswithcode

{ "category": "company", "query": "fintech startup payments Europe" }

Categories can be restrictive — try without one first if results are sparse.

Domain Filtering

Usually not needed. Use only when targeting specific sources:

{
  "includeDomains": ["arxiv.org", "github.com"],
  "excludeDomains": ["pinterest.com"]
}

Content Freshness

maxAgeHours controls cache vs livecrawl:

Value	Behavior
(omit)	Recommended default. Cache with livecrawl fallback
24	Livecrawl if cache > 24h old
0	Always livecrawl (real-time data)
-1	Cache only (fastest, for static/historical content)

Decision Tree

User needs web info
├─ Already has URLs? → /contents
├─ Wants a direct answer with citations? → /answer
└─ Needs to find pages?
   └─ /search
      ├─ Quick lookup → type: "fast"
      ├─ General search → type: "auto" (default)
      ├─ Deep research or structured data → type: "deep"
      └─ Complex multi-step reasoning → type: "deep-reasoning"

Common Parameter Mistakes — AVOID

• useAutoprompt → deprecated, do not use
• includeUrls / excludeUrls → do not exist. Use includeDomains / excludeDomains
• stream: true → not supported
• text/highlights at top level of /search → must be inside contents
• text/highlights inside contents on /contents → must be at top level
• livecrawl → deprecated. Use maxAgeHours instead
• numSentences, highlightsPerUrl → deprecated. Use max_characters

Output Formatting

After executing a search, present results to the user:

1. Summarize key findings in clear, concise French
2. Include source URLs for attribution
3. If using structured outputs, present the data in a readable format (table, list)
4. Flag low-confidence results if grounding data shows it

Always pipe curl through jq when available for readable output:

curl -s ... | jq .