web-crawler

Preferred entry: call exports.py (don't hand-roll requests)

Ready-made helpers live in skills/web-crawler/exports.py. Prefer them over writing your own proxied_get/proxied_post calls — they already inject the proxy credentials, so there is no API key to find (don't read $SCRAPECREATORS_API_KEY / $FIRECRAWL_API_KEY, don't check .env, don't ask the user).

import sys; sys.path.insert(0, "/data/workspace/skills/web-crawler")
from exports import scrape_markdown, youtube_transcript, sc_get

Named wrappers exist for the high-frequency actions (YouTube/TikTok transcript & video, IG/Twitter/Reddit posts, profiles, Google/Reddit search). For any other ScrapeCreators endpoint use sc_get(path, **params) — it auto-strips leading @/# from handles/hashtags. The intent-routing tables below still tell you which endpoint to pass. Pass caller_id="chat:<thread>" (or job:/preview:) for cost tracking.

Quick trigger rules (read this first)

Use this skill immediately when any of these conditions is true:

• web_fetch returns HTTP 401/403/429/5xx
• Response is an anti-bot challenge page (for example Cloudflare "Attention Required", "Just a moment", or challenge/captcha pages)
• The page is JS-heavy and the first fetch misses required detail fields (for example publish time, author, listing code, updated time, price breakdown)
• Search results do not contain the requested field and the value must be extracted from the target page itself

Fallback rule:

• If ordinary fetch is blocked or incomplete, switch to Firecrawl fallback in this skill before asking the user for screenshots/manual text.

Error signatures -> action

Paywall / Firecrawl-blocked fallback chain (use archive_fallback)

When Firecrawl can't get the page either (it returns 403, or markdown comes back empty) the site is behind a hard paywall or aggressive WAF. Do NOT keep retrying Firecrawl. Recover the article from a web archive snapshot instead:

How it works (and why this order):

• archive.today first (archive.ph / archive.is mirrors). User-triggered, real-browser captures; historically preserves full text behind paywalls. Best bet for NYT/WSJ/Economist. We scrape its /newest/ snapshot via Firecrawl (archive.today has its own Cloudflare, so scrape it through Firecrawl, never web_fetch it directly).
• Wayback Machine second (archive.org). Automated crawler that honors robots.txt and paywalls, so it often has NO full text for hard paywalls — but it's a good fallback for ordinary 403/Cloudflare pages that aren't paywalled.

Limitation: archives only return text someone already saved. If res["markdown"] == "", no snapshot exists — stop, tell the user, and try the outlet's official API/RSS or a different source. Do not fabricate the article.

China-app structured data (use apify_run)

When you need structured data from a China app — Douyin video search, Xiaohongshu notes, Weibo posts, Bilibili videos, JD/Taobao product prices, 1688 wholesale listings, Xianyu second-hand, Dewu sneakers, etc. — Firecrawl and ScrapeCreators don't cover these platforms. Use the Apify Store fallback instead. Apify is a serverless scraper marketplace with hundreds of community-maintained actors that run real browsers + proxy pools against Chinese platforms.

Auth: No user-supplied key needed. sc-proxy injects the platform Apify token automatically. The Authorization: Bearer header can be any fake value — the proxy replaces it with the real token. Do NOT read $APIFY_TOKEN from env, do NOT check .env, do NOT ask the user for an Apify key.

When to use Apify (vs Firecrawl/ScrapeCreators):

• ✅ China apps: 抖音, 小红书, 微博, B站, 京东, 淘宝, 1688, 闲鱼, 得物, 携程, 知乎, 豆瓣, 雪球, 快手, 爱奇艺, 优酷
• ✅ Southeast Asia e-commerce: Shopee, Lazada, Temu
• ❌ Western social media (TikTok/Instagram/YouTube/X/Reddit) → use ScrapeCreators first (cheaper)
• ❌ Generic web page scraping → use Firecrawl first (cheaper)
• ❌ Hard paywall articles → use archive_fallback (Apify doesn't help here)

How to pick an actor: The reliable-actors catalog is at output/apify_china_reliable.json (sorted by 30-day success count). Pick the top actor for the target platform. A few common ones:

Two-step Xiaohongshu workflow (search → note detail): The search scraper (zen-studio~rednote-search-scraper) returns only a truncated desc (~60 chars). For the full post body, run a second call with sian.agency~xiaohongshu-rednote-scraper in noteDetail mode, passing the id and xsec_token from the search result row. This is the only reliable way to get full note text for price/lodge/itinerary details.

Usage:

Billing & cost control — read this before every Apify call:

Apify uses pay-per-event billing: (actor-start + result_count × per_result + add-ons) × 2 credits. The dominant factor is result count × per-result price, and each actor prices differently ($0.003–$0.007/result). Unknown actors default to the highest tier ($0.007). Full pricing table and estimation examples: reference/apify-pricing.md.

You MUST estimate cost before calling:

1. Look up the actor's per-result price (table in reference/apify-pricing.md; unlisted = $0.007).
2. Estimate result count from input params (maxResults, maxProducts, etc.). If no limit is set, assume 100+.
3. Calculate: result_count × per_result × 2 = estimated credits.
4. If estimate > 5 credits, tell the user the cost before proceeding.

Hard spending cap — proxy-enforced default + per-call override:

The proxy automatically injects maxTotalChargeUsd=$2.5 (≈ 5 credits) on every actor run that doesn't already specify one. Apify terminates the run when the budget is hit and returns whatever results were collected so far — no overcharging, no runaway costs.

apify_run() also passes max_charge_usd (same default $2.5), which takes precedence over the proxy default. This means:

• Default (no extra param): capped at ≈ 5 credits per call. Safe for routine searches, profile lookups, small-batch scraping.
• Need more data? Raise the cap explicitly: max_charge_usd=10 (≈ 20 credits). Use when the user explicitly wants a large dataset and you've already told them the estimated cost.
• First-time test: lower it: max_charge_usd=0.5 (≈ 1 credit), limit results to 5–10, verify output quality before scaling up.
• Disable cap entirely: max_charge_usd=None. Never do this unless the user explicitly asks for an uncapped run after being warned of the potential cost.

Small-batch test first: When using an actor for the first time, set max_charge_usd=0.5 (≈ 1 credit) and limit results to 5–10. Verify output quality before scaling up.

Error handling:

• 400 run-failed → bad input (wrong field name). Fetch the actor's input-schema page and fix.
• 401 → proxy misconfigured (should not happen). Report to user.
• Empty result [] → actor ran but found nothing. Try different keywords or another actor.
• Timeout → increase timeout param (default 180s). Some actors are slow.

Cost discipline: Apify actors are more expensive than Firecrawl/ScrapeCreators. Only use Apify when the cheaper options can't get the data (China apps, structured e-commerce fields). For a single web page, always try Firecrawl first.

What each service is for

ScrapeCreators — Social media data extraction (27+ platforms)

Use for any request involving social media profiles, posts, videos, comments, transcripts, search, ads, trending content, or engagement metrics. Covers TikTok, Instagram, YouTube, LinkedIn, Facebook, Twitter/X, Reddit, Threads, Bluesky, Pinterest, Snapchat, Twitch, Kick, Truth Social, TikTok Shop, Google search, and link-in-bio services (Linktree, Komi, Pillar, Linkbio, Linkme, Amazon Shop).

Base URL: https://api.scrapecreators.com Auth: No user-supplied key needed. sc-proxy injects platform credentials automatically — just send the request. The x-api-key header can be any value or omitted entirely. Do NOT bail out or ask the user for a key if $SCRAPECREATORS_API_KEY looks unset; that env var is intentionally not required. Method: All endpoints use GET requests with query params. Responses are JSON.

Firecrawl — Fallback web page scraper

Only a fallback crawler for one web page when ordinary fetching fails. Use POST /v2/scrape with a single url and focused formats like markdown, html, rawHtml, links, summary, or constrained json/question/highlights extraction.

Auth: No user-supplied key needed. sc-proxy injects the Firecrawl credential automatically when you call through core.http_client.proxied_post — just send the request. Do NOT read $FIRECRAWL_API_KEY from env, do NOT check .env for it, and do NOT ask the user for a Firecrawl key if it looks unset; that env var is intentionally not required. The same proxy-injection model as ScrapeCreators applies here.

Do not use Firecrawl crawl/map/search/agent/browser endpoints. Do not request screenshots, audio, branding, images, or browser actions unless the proxy policy is expanded later.

ScrapeCreators — Intent routing

Map user intent to the right endpoint. Endpoint paths use the pattern /v1/platform/action.

Important: After selecting an endpoint from the tables below, fetch its OpenAPI spec at https://docs.scrapecreators.com/{path}/openapi.json for full parameter details, types, and example response before making the actual API call. For example: https://docs.scrapecreators.com/v1/tiktok/profile/openapi.json

Profiles / User Info

Posts / Content Feeds

Single Post / Video Details

Comments

Transcripts

⚠️ Metadata first. A transcript endpoint returns speech text only — no speaker, no title, no uploader. When summarizing a video, call the video metadata endpoint FIRST (youtube_video / /v1/tiktok/video / …) for title, description and uploader, then the transcript. Never attribute a speaker or public figure from transcript content alone — if metadata is unavailable, say "speaker unidentified" and summarize without naming anyone.

Search

Ad Libraries

Trending / Popular

Followers / Following / Live (TikTok only)

TikTok Shop

Link-in-Bio / Other

ScrapeCreators pagination

Paginated endpoints return a cursor/token in the response. Pass it back as a query param to get the next page.

ScrapeCreators known limitations

• Handles: pass without the @ symbol. Use charlidamelio not @charlidamelio. Applies to TikTok, Instagram, Twitter, Threads, Bluesky, Snapchat, Twitch, Pinterest, Truth Social
• YouTube handles: pass without the @ symbol. Use ThePatMcAfeeShow not @ThePatMcAfeeShow. You can also pass a channelId or full URL instead
• Hashtags: pass without the # symbol. Use fyp not #fyp. Applies to TikTok and YouTube hashtag search endpoints
• Twitter: returns ~100 most popular tweets, not chronological/latest
• Threads: only last 20-30 posts visible publicly
• Facebook posts: only 3 posts per page (API limitation)
• Facebook group posts: only 3 posts per page (same limitation)
• LinkedIn company posts: max 7 pages
• Instagram play counts: IG-only views (excludes cross-posted FB views)
• Truth Social: only prominent users (Trump, Vance, etc.) work publicly
• Transcripts: all transcript endpoints require video under 2 minutes
• Reddit subreddit names: case-sensitive! Use "AskReddit" not "askreddit"

Access patterns

ScrapeCreators (social media)

Use Python with core.http_client.proxied_get so sc-proxy injects credentials and bills correctly. Include a typed SC-CALLER-ID header (chat:, job:, preview:, etc.) for cost tracking. Do not read $SCRAPECREATORS_API_KEY from env and do not ask the user for a key — the proxy handles it.

Bash/curl works too (proxy is transparent), but Python is preferred for cost tracking:

Each endpoint has its own OpenAPI spec at https://docs.scrapecreators.com/{path}/openapi.json. Always fetch the per-endpoint spec first to get full parameter details before making the actual API call. The full spec is at https://docs.scrapecreators.com/openapi.json (large file — prefer per-endpoint specs).

Common optional params:

• trim (boolean): reduces response payload size. Use when you only need key metrics.
• region (string): 2-letter country code for proxy location. Does NOT filter by region — just routes through that country's proxy.

Firecrawl (web page fallback via transparent proxy)

No Firecrawl API key required — sc-proxy injects it. Don't look it up in .env or ask the user. Just call:

Decision rules

Route every request to the right backend. The user should never need to specify which API to use.

Social media request (profile, posts, comments, search, ads, trending, transcripts)

Use ScrapeCreators. Match the user's intent to an endpoint from the routing tables above. Strip @ from handles and # from hashtags before calling. Fetch the per-endpoint OpenAPI spec first for full param details.

YouTube URL or YouTube content request

Use ScrapeCreators for YouTube (channel info, videos, shorts, playlists, comments, transcripts, search, trending shorts). Match the user's intent to the appropriate YouTube endpoint from the routing tables above.

For a YouTube URL, use /v1/youtube/video to get video details and transcript. If the user's goal is content analysis, summarization, quote extraction, or topic mining, the transcript is included in the video response.

For a YouTube topic query, use /v1/youtube/search to find relevant videos, then call /v1/youtube/video only for the videos needed. Avoid fetching many videos by default.

Blocked or JS-heavy web page

Use Firecrawl once with formats:["markdown","links"] and onlyMainContent:true. Treat the returned Markdown as the extraction substrate, not as final truth: parse the title, price/value fields, specs, body description, image URLs, outbound links, and obvious contact/location hints from the page structure.

General web-page extraction lessons:

• Many listing/detail pages render important content with JavaScript, image galleries, hidden sections, or repeated UI labels. web_fetch may return boilerplate while Firecrawl can still recover the real main content.
• Do not hard-code site-specific labels. Convert page text into a generic structured summary: what it is, where it is, key numbers, evidence snippets, media/links, and caveats.
• Preserve source URLs for images and links when they help verify the page, but do not download or batch-process every media asset unless the user asks.
• If Markdown misses important layout or structured fields, retry once with rawHtml; use json, question, or highlights only when the user asked for narrow extraction and the schema/prompt is specific.

Cost discipline

ScrapeCreators — most endpoints cost 1 credit per request. Exceptions: /v1/tiktok/user/audience costs 26 credits; /v1/tiktok/video/transcript with use_ai_as_fallback=true costs +10 credits; /v1/google/company/ads with get_ad_details=true costs 25 credits. Warn users before calling expensive endpoints.

Firecrawl — billed per page plus expensive modifiers.

Keep calls tight: one page, one video, or a small shortlist. Never batch-crawl whole websites or bulk-scrape entire feeds with this skill.

If the proxy returns 403, the request is outside the allowed use case. Change the approach instead of retrying.

If the proxy returns 429, back off; do not parallelize around the limit.

If the upstream returns a failure, report the exact failure and avoid repeated paid retries unless one parameter change is clearly justified.**