smart-web

Execute an explicit bounded action batch in a task-owned tab of the approved local browser. Default: new tab closed after the batch. Opt-in session preserves top-level read/action/read state for at most 60 seconds and 10 calls in this owning server; resume with its opaque handle and exact current page URL, without navigation or replay. Use for user-authorized search/filter/expand workflows after retrieval alone is insufficient. Verify final evidence; a successful click is not goal completion. Password/MFA/CAPTCHA/consent require the user. Purchases, posting, deletion, uploads and account changes require separate explicit user authorization. No remote browser or credential export.

Search the web when the user gave a topic, keywords, or a `site:` query instead of a specific URL. Use this for discovery first; if you already have a direct URL, use smartfetch instead. Can optionally hand off into smartcrawl for same-site exploration. If smart-web returns reproducible incorrect or misleading output, open a GitHub issue on jojo-labs/smart-web with the exact input, arguments, observed output, expected behavior, and version. Skip obvious transient network/auth/rate-limit failures unless smart-web misclassified them.

Traverse a known site when one page is not enough and you need multiple relevant docs pages, posts, or thread items. Start from a concrete site URL; if you need to discover the site first, use smartsearch, and if you only need one direct URL, use smartfetch. Add `output_dir` when you want the crawl written to a local Markdown folder instead of returned only as a ranked summary; successful exports embed `resource_handle_manifest` in structuredContent with file:// handles for exported documents and metadata files. For LinkedIn profile URLs this is best-effort recent public post discovery, not full archive pagination. If smart-web returns reproducible incorrect or misleading output, open a GitHub issue on jojo-labs/smart-web with the exact input, arguments, observed output, expected behavior, and version. Skip obvious transient network/auth/rate-limit failures unless smart-web misclassified them.

Fetch and normalize one known URL or shortlink. Use this first for direct user-supplied URLs, browser-rendered pages, social posts, and videos; YouTube watch/short/embed links get best-effort transcript enrichment when public captions are available. For a blocked commerce URL, pass likely product/title candidates from the conversation as reference_queries; catalog metadata is accepted only after exact URL identifier matching. smartfetch is the unified retrieval kernel: default structuredContent carries pipeline acquire attempt_map, assessment, and research_handoff metadata for downstream evidence/report consumers when results are partial or blocked. Do not use smartsearch when the user already provided the URL, and use smartcrawl only when you need multiple pages from the same site. Use exactly one budget control per call: output_budget, preferred_budget_chars, context_mode, or headroom_tokens. If smart-web returns reproducible incorrect or misleading output, open a GitHub issue on jojo-labs/smart-web with the exact input, arguments, observed output, expected behavior, and version. Skip obvious transient network/auth/rate-limit failures unless smart-web misclassified them.

Read the selected smartvertical domain's typed input schema, supported actions and aliases before an unfamiliar call. No provider, browser or network access.

Korean/local vertical and structured data router. Use domain to access places, weather, subway, finance, DART, real estate, safety, news, commerce, spellcheck, zipcode, law, public restrooms, and GeekNews without exposing every provider as a separate MCP tool. For an unfamiliar domain, first read smartvertical_schema; unsupported domain fields, actions, sorts and incomplete coordinate pairs fail before provider I/O.