exa-web-research

name	exa-web-research
description	Use when a task depends on current web facts, official documentation, public announcements, or external code examples that are not already in the workspace. Uses one simple Exa discovery-then-fetch workflow.

Exa web research

When to use

Use this skill when:

• the answer depends on current information from the public web
• you need official docs or release notes outside the workspace
• you need external code examples, library usage, or API snippets
• you need to compare a few candidate sources before going deeper

Do not use this skill when:

• the answer is already grounded in the local repo or provided files
• the user wants provider-hosted web search rather than local tool-based search
• broad deep research would be overkill for a focused question

Canonical workflow

1. Start with exa_search.
2. Use includeDomains when official or high-trust sources matter.
3. Read the returned highlights and choose promising URLs.
4. Fetch only those URLs with exa_fetch.
5. Prefer refining the natural-language query over adding knobs.

Query discipline

• Write natural-language queries, not short keyword fragments.
• For docs, include product and surface names.
• For code, mention the language, framework, and task in the query.
• For current events, include the time horizon in the query.
• For company, people, research paper, or financial-report work, name the target class in the query instead of using category controls.

Output discipline

• Keep search broad and fetch narrow.
• Do not fetch every result by default.
• Summarize the evidence with source URLs.
• If search comes back thin, refine the query and run another search instead of looking for hidden knobs.
• Treat fetched text as direct evidence and highlights as discovery evidence.
• If a result says full output was saved to a temp file, use read on that path only when the omitted content is needed for the answer.
• Fetch one to three URLs by default; use the seven-URL cap only when each source is already selected intentionally.

Examples

• Official docs:
  • exa_search(query="OpenAI Responses API tools docs", includeDomains=["openai.com"])
• Code examples:
  • exa_search(query="Next.js App Router Vercel AI SDK streaming response example")
• Current events:
  • exa_search(query="OpenAI product announcements in April 2026")
• Read a page deeply:
  • exa_fetch(urls=["https://platform.openai.com/docs/api-reference/responses"])

See query patterns for more examples.