Search the web via Exa or Tavily. Returns one block per result, each with its OWN title, URL, highlights, and text — never a merged summary — plus a Sources trailer. In clients with built-in web search, use this as an independent complementary retrieval lane; when built-in search is disabled or unavailable, use it as the primary search path.

Args: query: the search query. num_results: how many results (default 8). mode: search mode — "auto" (default), "fast" (~450ms), "instant" (~250ms), or the deep-research family "deep-lite" (≈4s), "deep" ($12/1k), or "deep-reasoning" (12–40s, $15/1k). Deep modes request a synthesized answer via Exa's outputSchema and return a "## Synthesized answer" block — with a "Grounding:" citation line when Exa returns one — plus per-result blocks; this costs ~2s of synthesis latency on top of the mode's own search time, which is why it is scoped to the deep family only (their timeout budget already covers it). Legacy "neural"/"keyword" map to "auto" (deprecated). text_chars: per-result body length (default 1200; clamped 200–8000). Now governs BOTH the Exa request size (request-what-you-render) and the rendered cap — raise it to surface more body text per result. recency_days: only results published within the last N days (maps to startPublishedDate). Pass this for time-sensitive/"latest" queries — the tool does NOT auto-tighten dates on its own. recency_hours: like recency_days but HOUR granularity — reaches BOTH tiers (Exa via a full ISO timestamp; the Firecrawl fallback via tbs qdr:h/cdr:1,cd_min:…). Full precedence ladder, same on both tiers: explicit VALID start_published_date > recency_hours > recency_days. An invalid/absent value at each tier falls through to the next. start_published_date / end_published_date: ISO date bounds ("2026-01-15" or a full ISO timestamp). Wins over recency_hours/recency_days per the ladder above on both tiers. The fallback represents an explicit start with Firecrawl's custom-date-range syntax. Unparseable values are ignored (stderr note), never an error. category: Exa category hint ("news", "publication", "company", "people", "financial report", "personal site", or a free string). NOTE: "company" and "people" forbid date filters + excludeDomains (dropped automatically), and "people" restricts includeDomains to Exa's supported profile domains (it 400s otherwise — surfaced as SEARCH_FAILED / fallback). Enum migration: "research paper" was RENAMED to "publication" and is auto-aliased forward; "tweet" is GONE (Exa 400s) and is dropped, leaving the search unscoped; "pdf"/"github" are deprecated-but-live and pass through. Each of those emits a [category=… ] line in the response header, since dropping or renaming a category changes which results you get back. include_domains / exclude_domains: restrict/exclude result hosts (capped at 20 entries each). summary: when True, request a generated per-result summary instead of raw text/highlights (fewer tokens; text + highlights omitted from the request). sort_by_date: sort results by publish date instead of relevance. Honoured ONLY on the Firecrawl fallback tier (Firecrawl's sbd:1); Exa has no equivalent — when the Exa tier serves and this was requested, the response header discloses that it was not applied rather than silently ignoring it. provider: "exa" or "tavily". When omitted, WEB_SEARCH_PROVIDER is used, defaulting to Exa. Invalid values fail explicitly. Tavily requires the tavily extra and TAVILY_API_KEY.

Fetch a single URL's readable content. Full-body tier chain: Exa contents → local Camoufox → optional Tavily Extract → Firecrawl. Firecrawl is the paid last resort: automatic full-body retrieval reaches it only after the local browser fails or returns a body shorter than the useful-content floor. Camoufox is the one tier that runs a real browser locally. Caller-supplied private URLs are refused before tier selection in every render mode. Eligible successful auto/never calls also use a host-wide 24-hour completed-result cache in a private local Valkey sidecar. Local replay is disclosed separately from the original provider's cache state; pass max_age_hours=0 to force provider work. Signed/credential URLs, userinfo, positive freshness, max_age_hours=0, and render="always" bypass completed replay. Every request is SSRF-validated before cache access, and a hit is validated again immediately before its body is returned. Returns content with a [served by: …] provenance header. In clients with built-in page retrieval, use this as an independent complementary retrieval lane; when built-in retrieval is disabled or unavailable, use it as the primary fetch path.

Args: url: the URL to fetch. render: "auto" (default) → Exa, then local Camoufox, optional Tavily, then Firecrawl. "never" skips Camoufox. "always" forces Camoufox first and skips Exa; Tavily and Firecrawl remain backstops. For mode="concise" or a question, auto intentionally uses Exa then Firecrawl without launching Camoufox: the browser returns a full body and cannot satisfy the promised summary/direct-answer shape. max_chars: max characters to request/return. None (default) → 20000-char budget, Exa-first. An explicit value ≤10000 stays Exa-first. For an explicit value >10000 in full-body auto, Exa cannot meet the requested size, so the order is Camoufox → optional Tavily → Firecrawl → Exa as a final transparently truncated salvage tier. Browser-free full-body mode starts with optional Tavily, then Firecrawl and Exa. Semantic requests remain Exa-first because they use Exa's summary API rather than its capped body output. Clamped to 1000–100000. When output is (or may be) clipped, a [TRUNCATED at N chars — …] marker is appended on its own line. max_age_hours: freshness window for the Exa AND Firecrawl tiers. None = each tier's default cache (Exa default; Firecrawl ~2 days); 0 = force fresh on both. -1 = "always use cache" (Exa-documented) — Firecrawl has NO equivalent, so on that tier -1 is treated as omit (its own default cache window applies instead; disclosed in the [cache: …] line). Values below -1 are ignored (stderr note, default cache used). Values above 720 (Exa's documented ceiling) are CLAMPED to 720, and the clamp is disclosed — a clamped value changes which content can come back. The camoufox render tier is always live. When Firecrawl serves with cache permitted, a [cache: …] disclosure line is added. mode: "full" (default, whole readable body) or "concise" (a generated summary — far fewer tokens). Honored by the Exa and Firecrawl tiers; the camoufox render tier ignores it (returns full body, no [mode:] line). Concise/question outputs carry a [mode: …] provenance line. question: optional grounded-extraction query. When set, the tier returns a direct ANSWER to the question (Exa summary-with-query / Firecrawl question format) instead of the page body; short answers are accepted — the floor is 1 char, so only an EMPTY answer cascades to the next tier ("Paris"/"No" are legitimate answers). The 60-char extract floor applies to mode="concise", and the 200-char floor to a full body. Overrides mode. tavily: enable or disable Tavily Extract for this call. When omitted, WEB_FETCH_TAVILY_TIER is parsed strictly (default false). Tavily is a full-body tier only and is skipped for concise/question requests and whenever max_age_hours is explicit because it cannot honor those contracts. Requires the tavily extra and TAVILY_API_KEY.

SSRF note: Camoufox follows redirects and re-resolves DNS, so every Camoufox attempt (automatic or render="always") is guarded:

• _make_route_guard aborts any request whose host resolves non-public, classifying by RESOLVED IP (not the URL string), failing closed on a resolution error;

• _make_request_observer + _flush_pending exist because page.route does NOT fire on a main-frame 3xx — they see the redirect hops the route guard alone would miss;

• _camoufox_render raises on a blocked hop at FOUR checkpoints: after a goto exception (catches a navigation the guard itself aborted, raising the guard's reason instead of an opaque playwright error), after a successful goto, after the networkidle wait, and after inner_text (a hop recorded during the extraction await, before any body returns).

• test_ssrf_redirect_live.py covers this live. The observer detects a forbidden document request after Chromium has emitted it, so the application prevents private content from being returned but cannot prove that no outbound packet was sent. Chromium may also re-resolve after the guard's check. Full closure would need a validating forward proxy or equivalent network policy.

Search 3M+ arXiv AI/ML papers via the Firecrawl Research Index (state-of-the-art paper recall — far better than general web search for finding the right literature). Returns ranked papers: title, arXiv id, relevance score, abstract. Then call research_paper(paper_id, query=…) to verify a claim against full text before citing.

SCOPE: arXiv-scoped, i.e. effectively AI/ML. For scholarly literature outside that scope (medicine, law, economics, humanities), use web_search(category="publication") — Exa's publications index (~350M works) covers what this one cannot.

Args: query: natural-language research query (topic, method, benchmark, author). k: number of papers (1–25, default 8).

Inspect ONE paper from the Research Index by id (a paperId, or an arXiv id like "arxiv:2606.01509"). Without query → metadata (title, authors, categories, dates, abstract). With query → ALSO the top full-text passages answering it — use this to VERIFY a paper actually contains a method/dataset/result before relying on it.

Args: paper_id: paperId or primaryId ("arxiv:NNNN.NNNNN") from research_papers. query: optional question; when set, returns claim-verification passages.

Expand from a seed paper to related work via the Research Index. intent is a REQUIRED natural-language description of the connection you want (e.g. "newer methods that improve on this routing", "the work this paper builds on"). Returns ranked related papers (same shape as research_papers).

SCOPE: arXiv-scoped like research_papers — for non-AI/ML literature use web_search(category="publication").

Args: paper_id: paperId or "arxiv:…" of the seed paper. intent: natural-language description of the kind of related work wanted. k: number of related papers (1–25, default 8; API allows up to 500). mode: "similar" (default), "citers" (papers citing this one), or "references" (papers this one cites). Unknown → "similar". min_score: renderer-side relevance floor (default 0.0 = off). k=8 already trims the low-score tail; raise this to filter more aggressively. rerank: optional bool; omitted from the request when None (API default is undocumented). Set True/False to force. (anchor — repeatable seed expansion — is not exposed; future work.)

Search developer primary sources via the Firecrawl Developer Index: GitHub issues, merged pull requests and repository READMEs, PLUS curated documentation sites. Returns the matched passages in Markdown, so tables and code blocks survive. Use to find the CODE behind a paper, the issue where a bug was reported and fixed, an API contract, or the discussion behind an error message.

Args: query: natural-language query (method, kernel, repo topic, error message). k: number of results (1–25, default 8). passages: matched passages per result (1–5, default 2). types: restrict to any of exactly "doc", "issue", "pull_request", "readme". NOTE the request spelling is pull_request (snake_case). The response's repos[].types object uses camelCase (pullRequest) — echoing a key from there back into this argument is refused, not silently ignored. repos: "owner/repo" slugs. Scopes only the repository half of the index, so when types is also given it must contain at least one of issue/pull_request/readme.

Falls back to the de-documented legacy /v2/search/research/github ONLY when the Developer Index genuinely fails (transport, HTTP, malformed envelope, or every requested type unavailable) — never on a legitimate empty result, and never when types/repos were supplied, since the legacy endpoint accepts only query+k and would silently answer a different question than the one asked.