<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Arco Labs</title>
    <description>The latest articles on DEV Community by Arco Labs (@arcolabs).</description>
    <link>https://dev.to/arcolabs</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4250875%2F5ba8771e-f5f2-40c0-8d7c-200d21d6d124.png</url>
      <title>DEV Community: Arco Labs</title>
      <link>https://dev.to/arcolabs</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/arcolabs"/>
    <language>en</language>
    <item>
      <title>Why Blindly Feeding MP4s to LLMs Fails (and How Progressive Video Disclosure Fixes It)</title>
      <dc:creator>Arco Labs</dc:creator>
      <pubDate>Wed, 18 Nov 2026 03:27:11 +0000</pubDate>
      <link>https://dev.to/arcolabs/why-blindly-feeding-mp4s-to-llms-fails-and-how-progressive-video-disclosure-fixes-it-5544</link>
      <guid>https://dev.to/arcolabs/why-blindly-feeding-mp4s-to-llms-fails-and-how-progressive-video-disclosure-fixes-it-5544</guid>
      <description>&lt;p&gt;If you give a coding or research agent a YouTube or TikTok URL today, the naive pipeline looks straightforward: run &lt;code&gt;yt-dlp&lt;/code&gt;, download the MP4, and pass the video bytes to a vision-capable model.&lt;/p&gt;

&lt;p&gt;In production, that pipeline breaks for four separate reasons before your agent even reads a single sentence:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Most frontier coding agents cannot take raw video files at all.&lt;/strong&gt; Anthropic Claude accepts images (up to 600 per request within a 128 MB cap) but has no native video or audio input. OpenAI's Responses API similarly takes images and documents rather than arbitrary social video URLs. Even when a model family accepts video natively, you still need a deterministic, cacheable text-and-frame layer if your pipeline feeds Claude, GPT, DeepSeek, or a RAG vector store.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Token and bandwidth costs scale linearly with video length.&lt;/strong&gt; On a 30-minute video at high visual resolution, dense frame sampling burns hundreds of thousands of tokens when 90% of the answer sits in the spoken transcript and three specific slides.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Platform CDN links are IP-bound and short-lived.&lt;/strong&gt; Passing a raw &lt;code&gt;googlevideo.com&lt;/code&gt; or &lt;code&gt;tiktokcdn.com&lt;/code&gt; URL straight to an external LLM API returns &lt;code&gt;403 Forbidden&lt;/code&gt; because the media signature is tied to the IP address that negotiated the player session.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Datacenter IPs hit silent bot traps.&lt;/strong&gt; If you deploy an extraction service onto cloud infrastructure such as Hetzner, Vultr, or AWS, YouTube responds to most real videos with &lt;code&gt;LOGIN_REQUIRED&lt;/code&gt; while letting famous test fixtures succeed.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;We built &lt;a href="https://vidcrawl.com" rel="noopener noreferrer"&gt;VidCrawl&lt;/a&gt; around &lt;strong&gt;progressive video disclosure&lt;/strong&gt;: keep Layer 1 purely deterministic (zero LLM tokens), send lightweight storyboard grids alongside timestamped text, and fetch high-resolution pixels only at the exact timestamps where text alone leaves ambiguity.&lt;/p&gt;

&lt;p&gt;Here is how the extraction engine works under the hood, including the production failure modes we hit along the way.&lt;/p&gt;




&lt;h2&gt;
  
  
  Trap #1: The Hetzner &lt;code&gt;LOGIN_REQUIRED&lt;/code&gt; and Rickroll (&lt;code&gt;dQw4w9WgXcQ&lt;/code&gt;) Whitelist Trap
&lt;/h2&gt;

&lt;p&gt;Early in testing from our Hetzner (&lt;code&gt;AS24940&lt;/code&gt;) server, our YouTube smoke tests passed against &lt;code&gt;dQw4w9WgXcQ&lt;/code&gt;, and then failed immediately when real users pasted real videos like &lt;code&gt;1E3tv_3D95g&lt;/code&gt;, &lt;code&gt;UF8uR6Z6KLc&lt;/code&gt;, or &lt;code&gt;aircAruvnKk&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;From flagged cloud ASNs, YouTube's &lt;code&gt;/youtubei/v1/player&lt;/code&gt; and watch HTML return:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;playabilityStatus.status = "LOGIN_REQUIRED"
reason = "Sign in to confirm you're not a bot"
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Meanwhile, &lt;code&gt;dQw4w9WgXcQ&lt;/code&gt; continues returning &lt;code&gt;OK&lt;/code&gt; from the exact same datacenter IP. If your CI or &lt;code&gt;/healthz&lt;/code&gt; check tests &lt;code&gt;dQw4w9WgXcQ&lt;/code&gt;, your monitoring stays green while 90%+ of real production URLs fail.&lt;/p&gt;

&lt;p&gt;A second mistake is masking that failure: returning &lt;code&gt;ok: true&lt;/code&gt; with zero transcript segments when &lt;code&gt;playabilityStatus.status === "LOGIN_REQUIRED"&lt;/code&gt;. VidCrawl excludes &lt;code&gt;dQw4w9WgXcQ&lt;/code&gt; from live reliability suites (testing non-whitelisted URLs such as &lt;code&gt;1E3tv_3D95g&lt;/code&gt; instead) and raises a typed &lt;code&gt;YOUTUBE_BOT_CHALLENGED&lt;/code&gt; (&lt;code&gt;502&lt;/code&gt;) error whenever &lt;code&gt;/player&lt;/code&gt; returns &lt;code&gt;LOGIN_REQUIRED&lt;/code&gt; rather than returning an empty success payload.&lt;/p&gt;




&lt;h2&gt;
  
  
  Fix #1: Split-Routing &lt;code&gt;/next&lt;/code&gt;, Sticky Proxy &lt;code&gt;/player&lt;/code&gt;, Direct &lt;code&gt;json3&lt;/code&gt;, and &lt;code&gt;i.ytimg.com&lt;/code&gt;
&lt;/h2&gt;

&lt;p&gt;Routing an entire video download through a residential proxy destroys unit economics. Instead, VidCrawl splits every YouTube scrape across four distinct network paths in &lt;code&gt;packages/engine/src/youtube.ts&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;POST /v1/scrape (1 credit, 0 LLM tokens)
  |
  +---&amp;gt; 1. Direct server IP: POST /youtubei/v1/next (WEB client)
  |        -&amp;gt; visitorData, attributedDescription, outgoing links, macro-marker chapters
  |        -&amp;gt; 0 proxy bytes (skips ~1.2 MB watch HTML)
  |
  +---&amp;gt; 2. Sticky residential proxy: POST /youtubei/v1/player
  |        -&amp;gt; Cycles VISIONOS / ANDROID_VR / IOS clients on deterministic sticky port
  |        -&amp;gt; ~20-40 KB compressed JSON per call
  |
  +---&amp;gt; 3. Direct server IP: GET i.ytimg.com/sb/... (highest available L3/L2 storyboards)
  |        -&amp;gt; Sprite sheets of 3x3 or 5x5 tile grids (0 proxy bytes)
  |
  +---&amp;gt; 4. Direct server IP: GET timedtext?fmt=json3
           -&amp;gt; Clean tStartMs + dDurationMs segments (0 proxy bytes)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Three implementation details matter here:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic sticky proxy ports (&lt;code&gt;getStickyProxyUrl&lt;/code&gt;)&lt;/strong&gt;: For DataImpulse (&lt;code&gt;gw.dataimpulse.com:823&lt;/code&gt;), ports &lt;code&gt;10000..19999&lt;/code&gt; map to sticky residential sessions. Hashing &lt;code&gt;${videoId}:${attempt}&lt;/code&gt; onto a sticky port keeps the &lt;code&gt;/player&lt;/code&gt; negotiation and any subsequent IP-bound media or audio fetch on the same exit IP while rotating across videos and retry attempts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;json3&lt;/code&gt; captions instead of naive WebVTT&lt;/strong&gt;: YouTube's auto-generated &lt;code&gt;.vtt&lt;/code&gt; tracks use a two-line rolling display window where every cue repeats the previous line, causing duplicated sentences in downstream transcripts. Fetching &lt;code&gt;&amp;amp;fmt=json3&lt;/code&gt; (with &lt;code&gt;srv3&lt;/code&gt; and XML fallbacks) yields discrete &lt;code&gt;events[].segs[].utf8&lt;/code&gt; entries keyed by &lt;code&gt;tStartMs&lt;/code&gt; and &lt;code&gt;dDurationMs&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Bun runtime requirement for &lt;code&gt;fetch({ proxy })&lt;/code&gt;&lt;/strong&gt;: Bun natively supports &lt;code&gt;fetch(url, { proxy })&lt;/code&gt;, whereas Node ignores &lt;code&gt;proxy&lt;/code&gt; on &lt;code&gt;RequestInit&lt;/code&gt; and falls back to direct egress. VidCrawl checks for Bun at runtime when &lt;code&gt;VIDCRAWL_PROXY_URL&lt;/code&gt; is configured so traffic never leaks directly by mistake.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;On our recorded 18:22 YouTube sample (&lt;code&gt;WWDC 2025 Impressions: Liquid Glass!&lt;/code&gt;, &lt;code&gt;1E3tv_3D95g&lt;/code&gt;), Layer 1 produces &lt;strong&gt;26 KB of Markdown&lt;/strong&gt; (465 transcript segments and 9 chapters) plus &lt;strong&gt;237 KB across 5 storyboard sheets&lt;/strong&gt; (112 tiles sampled every 10 seconds). An agent can read the entire talk and scan every visual transition before requesting a single full-resolution frame. If you want to test the Markdown output directly on a browser URL, see our &lt;a href="https://vidcrawl.com/video-to-markdown" rel="noopener noreferrer"&gt;Video to Markdown&lt;/a&gt; tool.&lt;/p&gt;




&lt;h2&gt;
  
  
  Fix #2: ISO-BMFF &lt;code&gt;sidx&lt;/code&gt; Byte-Range Frame Slicing (&lt;code&gt;POST /v1/frames&lt;/code&gt;)
&lt;/h2&gt;

&lt;p&gt;When a storyboard grid shows an unreadable 14px IDE buffer, chart axis, or settings screen, the agent calls &lt;code&gt;POST /v1/frames&lt;/code&gt; with &lt;code&gt;timestamps_sec&lt;/code&gt; (1 to 20 timestamps, &lt;code&gt;max_width&lt;/code&gt; from 320px to 1920px).&lt;/p&gt;

&lt;p&gt;To avoid downloading a multi-hundred-megabyte MP4 just to grab two screenshots, &lt;code&gt;packages/engine/src/youtube.ts&lt;/code&gt; caches the best MP4 DASH stream from &lt;code&gt;/player&lt;/code&gt; for 15 minutes, preferring 720p streams (&lt;code&gt;compareFrameStreams&lt;/code&gt; scores by short side so vertical 720x1280 Shorts do not stall on giant 1080x1920 segments) along with &lt;code&gt;initRange&lt;/code&gt; and &lt;code&gt;indexRange&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Then &lt;code&gt;packages/engine/src/frames.ts&lt;/code&gt; performs exact byte-range slicing:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Fetch the container header (&lt;code&gt;0..indexRange.end&lt;/code&gt;)&lt;/strong&gt; capped at &lt;code&gt;MAX_HEADER_BYTES = 2 MB&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Parse the ISO-BMFF &lt;code&gt;sidx&lt;/code&gt; (Segment Index Box, &lt;code&gt;0x73696478&lt;/code&gt;)&lt;/strong&gt; in TypeScript (&lt;code&gt;findSegmentByteRangeFromSidx&lt;/code&gt;), validating version 0 and version 1 reference entries and accumulating timestamps across &lt;code&gt;timescale&lt;/code&gt; ticks until &lt;code&gt;targetSec&lt;/code&gt; lands inside &lt;code&gt;[startSec, endSec)&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fetch only &lt;code&gt;initRange&lt;/code&gt; plus that single 5-second GOP fragment &lt;code&gt;[byteStart, byteEnd]&lt;/code&gt;&lt;/strong&gt; capped at &lt;code&gt;MAX_MEDIA_BYTES = 16 MB&lt;/code&gt;, enforcing manual redirects (&lt;code&gt;redirect: "manual"&lt;/code&gt;) on every hop so redirects can never leave &lt;code&gt;googlevideo.com&lt;/code&gt;, and reusing the exact &lt;code&gt;stickyProxyUrl&lt;/code&gt; from &lt;code&gt;/player&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Decode in memory with &lt;code&gt;ffmpeg&lt;/code&gt; and &lt;code&gt;sharp&lt;/code&gt;&lt;/strong&gt;: Concatenate &lt;code&gt;[initBuf, segBuf]&lt;/code&gt;, pipe into &lt;code&gt;ffmpeg&lt;/code&gt; (&lt;code&gt;select=gte(n,${targetFrameIdx}) -frames:v 1&lt;/code&gt;), resize to &lt;code&gt;max_width&lt;/code&gt;, and encode as JPEG (&lt;code&gt;quality: 88&lt;/code&gt;).&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If the stream header or &lt;code&gt;sidx&lt;/code&gt; segment is unavailable, &lt;code&gt;/v1/frames&lt;/code&gt; falls back explicitly to &lt;code&gt;storyboard_crop&lt;/code&gt; by extracting the closest tile from the storyboard JPEG via &lt;code&gt;sharp&lt;/code&gt; and labeling the &lt;code&gt;source&lt;/code&gt; field as &lt;code&gt;"storyboard_crop"&lt;/code&gt; rather than &lt;code&gt;"ffmpeg_seek"&lt;/code&gt;.&lt;/p&gt;




&lt;h2&gt;
  
  
  Handling Short-Form Video (TikTok, Instagram Reels, X)
&lt;/h2&gt;

&lt;p&gt;Short-form clips do not expose YouTube-style &lt;code&gt;sidx&lt;/code&gt; byte ranges or static &lt;code&gt;i.ytimg.com&lt;/code&gt; storyboards, and they frequently deliver critical text as burned-in overlays rather than spoken words.&lt;/p&gt;

&lt;p&gt;For TikTok, VidCrawl uses a two-rung pipeline (&lt;code&gt;packages/engine/src/tiktok.ts&lt;/code&gt; and &lt;code&gt;short-form-media.ts&lt;/code&gt;) that powers both &lt;code&gt;/v1/scrape&lt;/code&gt; and our public &lt;a href="https://vidcrawl.com/tiktok-transcript" rel="noopener noreferrer"&gt;TikTok transcript&lt;/a&gt; tool:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Rung 1 (Zero-&lt;code&gt;yt-dlp&lt;/code&gt; Hydration JSON)&lt;/strong&gt;: Fetch the page HTML and parse &lt;code&gt;&amp;lt;script id="__UNIVERSAL_DATA_FOR_REHYDRATION__"&amp;gt;&lt;/code&gt; (&lt;code&gt;webapp.video-detail.itemInfo.itemStruct&lt;/code&gt;), pulling native ASR WebVTT tracks directly from &lt;code&gt;subtitleInfos&lt;/code&gt; when present.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rung 2 (Lowest-Bitrate H.264 + Direct-First Fallback)&lt;/strong&gt;: When building the &lt;code&gt;3x3&lt;/code&gt; contact sheet (&lt;code&gt;960x1704&lt;/code&gt;, &lt;code&gt;320x568&lt;/code&gt; tiles via &lt;code&gt;ffmpeg&lt;/code&gt;) or running Whisper ASR, select lowest-bitrate &lt;code&gt;h264&lt;/code&gt;/&lt;code&gt;avc&lt;/code&gt; streams rather than &lt;code&gt;bytevc1&lt;/code&gt; (which on TikTok frequently omits audio). Downloads run directly from the server IP first and fall back to the residential proxy only when the CDN answers HTTP 403 (&lt;code&gt;~17%&lt;/code&gt; of TikTok clips from production egress).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Instagram Reels (&lt;code&gt;embed/captioned/&lt;/code&gt; HTML + bounded clip processing) and X posts (&lt;code&gt;cdn.syndication.twimg.com/tweet-result&lt;/code&gt; + &lt;code&gt;video.twimg.com&lt;/code&gt; clips) follow the same bounded &lt;code&gt;50 MB&lt;/code&gt; / &lt;code&gt;600s&lt;/code&gt; pipeline and return the exact same JSON schema.&lt;/p&gt;




&lt;h2&gt;
  
  
  Measured Reliability on Hetzner Egress (300-URL Benchmark)
&lt;/h2&gt;

&lt;p&gt;Instead of guessing reliability, we run a deterministic 300-URL evaluation suite (&lt;code&gt;eval/reliability/run.ts&lt;/code&gt;, 60 public URLs per platform set, concurrency 3, 16.9 minutes total) from Hetzner egress (&lt;code&gt;eval/reliability/results/2026-11-16-hetzner-final.md&lt;/code&gt;):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Platform Set&lt;/th&gt;
&lt;th&gt;Scrape OK&lt;/th&gt;
&lt;th&gt;Timed Transcript&lt;/th&gt;
&lt;th&gt;Any Text&lt;/th&gt;
&lt;th&gt;Storyboard&lt;/th&gt;
&lt;th&gt;Frame OK (1280px)&lt;/th&gt;
&lt;th&gt;Scrape p50 / p95&lt;/th&gt;
&lt;th&gt;Proxy KB / Scrape&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;YouTube&lt;/td&gt;
&lt;td&gt;97% (58/60)&lt;/td&gt;
&lt;td&gt;97% (56/58)&lt;/td&gt;
&lt;td&gt;97% (56/58)&lt;/td&gt;
&lt;td&gt;100% (58/58)&lt;/td&gt;
&lt;td&gt;95% (19/20)&lt;/td&gt;
&lt;td&gt;2.4s / 10.4s&lt;/td&gt;
&lt;td&gt;189 KB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;YouTube Shorts&lt;/td&gt;
&lt;td&gt;98% (59/60)&lt;/td&gt;
&lt;td&gt;92% (54/59)&lt;/td&gt;
&lt;td&gt;92% (54/59)&lt;/td&gt;
&lt;td&gt;100% (59/59)&lt;/td&gt;
&lt;td&gt;84% (16/19)&lt;/td&gt;
&lt;td&gt;3.1s / 14.7s&lt;/td&gt;
&lt;td&gt;55 KB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;TikTok&lt;/td&gt;
&lt;td&gt;100% (60/60)&lt;/td&gt;
&lt;td&gt;95% (57/60)&lt;/td&gt;
&lt;td&gt;95% (57/60)&lt;/td&gt;
&lt;td&gt;88% (53/60)&lt;/td&gt;
&lt;td&gt;85% (17/20)&lt;/td&gt;
&lt;td&gt;7.7s / 24.1s&lt;/td&gt;
&lt;td&gt;229 KB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Instagram Reels&lt;/td&gt;
&lt;td&gt;100% (60/60)&lt;/td&gt;
&lt;td&gt;93% (56/60)&lt;/td&gt;
&lt;td&gt;93% (56/60)&lt;/td&gt;
&lt;td&gt;100% (60/60)&lt;/td&gt;
&lt;td&gt;100% (20/20)&lt;/td&gt;
&lt;td&gt;11.5s / 27.6s&lt;/td&gt;
&lt;td&gt;0 KB&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;X Video&lt;/td&gt;
&lt;td&gt;100% (60/60)&lt;/td&gt;
&lt;td&gt;82% (49/60)&lt;/td&gt;
&lt;td&gt;100% (60/60)&lt;/td&gt;
&lt;td&gt;100% (60/60)&lt;/td&gt;
&lt;td&gt;100% (20/20)&lt;/td&gt;
&lt;td&gt;2.7s / 8.7s&lt;/td&gt;
&lt;td&gt;0 KB&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Observed failure modes in that run were explicit and untruncated: 2 &lt;code&gt;VIDEO_UNAVAILABLE&lt;/code&gt; and 1 &lt;code&gt;fell_back_to_storyboard_crop&lt;/code&gt; on YouTube, 1 scrape &lt;code&gt;UPSTREAM_TIMEOUT&lt;/code&gt; and 3 frame &lt;code&gt;UPSTREAM_TIMEOUT&lt;/code&gt; on Shorts, and 3 &lt;code&gt;fell_back_to_storyboard_crop&lt;/code&gt; on TikTok. Every failed request before model completion is automatically excluded from billing.&lt;/p&gt;




&lt;h2&gt;
  
  
  Connecting via Remote HTTP MCP or &lt;code&gt;npx -y @vidcrawl/mcp&lt;/code&gt;
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://vidcrawl.com/mcp" rel="noopener noreferrer"&gt;VidCrawl MCP server&lt;/a&gt; exposes five tools that mirror the progressive-disclosure workflow:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;vidcrawl_scrape&lt;/code&gt; (1 credit): Metadata, chapters, transcript, storyboard sheets, and Markdown.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;vidcrawl_frames&lt;/code&gt; (1 credit): Up to 20 sharp JPEG frames at exact &lt;code&gt;timestamps_sec&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;vidcrawl_search&lt;/code&gt; (1 credit, +1 per scraped result): YouTube discovery with optional Layer 1 scrape enrichment.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;vidcrawl_agent&lt;/code&gt; (3 / 5 / 10 credits): Bounded visual + spoken synthesis across 1 to 10 URLs with &lt;code&gt;[MM:SS]&lt;/code&gt; citations and optional JSON schema.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;vidcrawl_account&lt;/code&gt; (0 credits): Check remaining credits and endpoint prices.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Option A: Remote Streamable HTTP (Zero Local Dependencies)
&lt;/h3&gt;

&lt;p&gt;In Claude Code:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude mcp add --transport http vidcrawl https://vidcrawl.com/mcp \
  --header &lt;span class="s2"&gt;"Authorization: Bearer $VIDCRAWL_API_KEY"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In OpenAI Codex:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;codex mcp add vidcrawl --url https://vidcrawl.com/mcp
&lt;span class="c"&gt;# In ~/.codex/config.toml under [mcp_servers.vidcrawl], add:&lt;/span&gt;
&lt;span class="c"&gt;# bearer_token_env_var = "VIDCRAWL_API_KEY"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In Cursor or Windsurf (&lt;code&gt;.cursor/mcp.json&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"vidcrawl"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"https://vidcrawl.com/mcp"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"headers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="nl"&gt;"Authorization"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Bearer YOUR_VIDCRAWL_API_KEY"&lt;/span&gt;
      &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Option B: Local Stdio MCP and CLI (&lt;code&gt;@vidcrawl/mcp&lt;/code&gt;)
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Run as a stdio MCP server&lt;/span&gt;
&lt;span class="nv"&gt;VIDCRAWL_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"$VIDCRAWL_API_KEY"&lt;/span&gt; &lt;span class="nv"&gt;VIDCRAWL_API_URL&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"https://vidcrawl.com"&lt;/span&gt; npx -y @vidcrawl/mcp mcp

&lt;span class="c"&gt;# Or call the CLI directly from a shell script&lt;/span&gt;
npx -y @vidcrawl/mcp scrape &lt;span class="s1"&gt;'https://www.youtube.com/watch?v=1E3tv_3D95g'&lt;/span&gt;
npx -y @vidcrawl/mcp frames &lt;span class="s1"&gt;'https://www.youtube.com/watch?v=1E3tv_3D95g'&lt;/span&gt; --at 150,365
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  Pricing and Alternatives
&lt;/h2&gt;

&lt;p&gt;New accounts include 100 free credits with no card required, and one-time credit packs start at &lt;code&gt;$9 / 1,000 credits&lt;/code&gt; and &lt;code&gt;$29 / 4,000 credits&lt;/code&gt; (credits never expire). Every &lt;code&gt;/v1/scrape&lt;/code&gt;, &lt;code&gt;/v1/frames&lt;/code&gt;, and &lt;code&gt;/v1/search&lt;/code&gt; call costs 1 credit regardless of whether a clip is 15 seconds or 2 hours long, and &lt;code&gt;/v1/agent&lt;/code&gt; costs 3, 5, or 10 credits depending on effort rather than per-minute billing. If you are evaluating alternatives on per-minute vs per-call economics, SDKs, and platform coverage, check our side-by-side breakdown on &lt;a href="https://vidcrawl.com/alternative/supadata" rel="noopener noreferrer"&gt;VidCrawl vs Supadata&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Try the live playground or create an API key at &lt;a href="https://vidcrawl.com" rel="noopener noreferrer"&gt;VidCrawl&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>webdev</category>
      <category>ai</category>
      <category>mcp</category>
      <category>typescript</category>
    </item>
  </channel>
</rss>
