| name | exa-web-search |
| description | Look something up on the web with Exa MCP — quick, source-backed search and page-fetch. Use when the user asks to search the web, look something up, find current/latest information, get a fact with a citation, fetch a specific page/doc, or look up a person/company/package. For a deep, multi-source, fact-checked investigation or report, use the deep-research skill instead. |
Exa Web Search
Use this skill for quick, source-backed lookups with the Exa MCP tools: a search, optionally a page fetch, and a concise sourced answer.
This is the single-query search skill. For a full multi-source investigation — fanning out many searches, cross-verifying claims across sources, and synthesizing a structured report — use the deep-research skill instead. If a request is clearly a deep investigation ("research X thoroughly", "compare A vs B across the field", "state-of-the-art survey of Y"), defer to deep-research.
When to Use
Use Exa search when the user asks to:
- search the web, look something up, or "google" something
- find latest/current/recent or time-sensitive information
- get a fact with a citation, or verify a claim beyond local files / model knowledge
- fetch the content of a specific page, doc, changelog, or release note
- look up a person, company, or package
Do not search automatically for tasks answerable from the local repository or already-provided context, unless the user asks for current/external information.
Preferred Tools
Prefer the discovered direct tools when available:
mcp_exa_web_search_exa — search the web
mcp_exa_web_fetch_exa — fetch full page content for selected URLs
If direct tools are unavailable, use mcp_list to inspect MCP servers and then mcp_call with the exa server.
Search Workflow
- Clarify the target if it is ambiguous.
- Search with
mcp_exa_web_search_exa using a semantic query that describes the ideal source/page.
- Review result titles, URLs, snippets/highlights, dates, and source quality.
- Fetch the most relevant primary source(s) with
mcp_exa_web_fetch_exa when snippets are insufficient or a citation needs verification.
- Answer concisely with source URLs, separating confirmed facts from inference.
Keep it light: a single search (plus maybe one or two fetches) is the norm. If you find yourself needing many searches and cross-checking across numerous sources, that's deep-research territory — say so and switch.
Query Guidelines
Write natural-language queries describing the ideal page, not keyword soup.
Good:
official Pi coding agent documentation MCP support directTools latest release notes
npm package page <package> latest version published date README
GitHub issues <org>/<repo> <topic>
Poor:
Pi MCP
directTools latest
For current information, include recency intent and likely source types: latest, current, 2026, release notes, changelog, documentation, npm package, GitHub issue, official docs.
For people/company searches, use Exa category prefixes:
category:people Jane Doe machine learning engineer LinkedIn
category:company Acme AI funding product leadership
Use a small result set first, usually numResults: 5 to 10. Increase only when coverage is poor.
Fetching Pages
Use mcp_exa_web_fetch_exa when:
- the answer depends on exact wording, dates, versions, or configuration details
- snippets conflict or are incomplete
- the user explicitly asks for sources/citations
- results point to docs, package registries, GitHub issues, or release pages worth quoting
Fetch 2–5 high-quality URLs rather than many low-quality pages. Prefer official/primary sources: official docs/vendor pages → package registries/release notes → GitHub repos/issues/PRs/changelogs → standards/spec pages → reputable technical blogs/news only when primary sources are absent.
Citation Rules
- Always cite URLs for web-derived claims.
- Do not invent citations or cite sources you did not inspect.
- If only search highlights were used, phrase carefully and include the result URL.
- For important claims, fetch the page before citing it as evidence.
- Mention dates or versions when relevant.
- If sources disagree, say so and explain which appears authoritative.
Answer Format
Keep answers tight — this is a lookup, not a report.
Summary sentence.
Sources:
- https://example.com/source-1
- https://example.com/source-2
Do not dump raw search results unless asked. If the answer is growing into a structured multi-section report, that's a sign the request was really deep-research.
Quality Checks
Before finalizing:
- Are the sources primary and recent enough?
- Did you fetch pages where snippets were insufficient?
- Are the non-obvious factual claims sourced?
- Did you avoid overclaiming beyond the evidence?
- Did you include exact versions/dates if the user asked for latest/current info?