Docs

start_video_analysis

Submit a video for AI vision analysis and return immediately with a job_id — do NOT wait for the analysis to finish in this call. Use when you want to run OTHER Algrow tool calls (channel data, viral search, etc.) in parallel while the analysis processes in the background. Pair with get_video_analysis_result(job_id) to pick up the finished analysis later. Same backend job as analyze_video: 1 credit per 4 minutes of video (3x at default resolution), max 3hr videos, 2hr YouTube cache applies.

"Start analyzing https://youtube.com/watch?v=abc123 for hook patterns, I'll come back for the result"

get_video_analysis_result

Fetch the result of a video analysis job submitted via start_video_analysis. Blocks until the job finishes (up to several minutes) and then renders the analysis widget. Always pass video_url and prompt back in — the widget needs them to render the embedded player and the prompt header.

"Get the finished video analysis for that job I started earlier"

download_video

Save a YouTube video, audio track, or subtitle file to a stable URL the user can click to download. Lands the bytes in Algrow’s storage and returns a deterministic mp4/mp3/srt URL with Content-Disposition attachment so clicking it triggers a save. Renders the dino runner game while the upstream download is running so users have something to do during the ~60–120s cold fetch.

"Save the audio track of that video as mp3"

"Grab the subtitles for the last video we analyzed"

"Save just the 1:30 to 2:00 section of this video in 1080p"

Formats. video returns mp4 at the requested quality (360p / 480p / 720p default / 1080p). audio returns mp3 (~10× smaller than video). subtitles returns SRT — tries human-authored first, then auto-generated English.

Timestamps. start and end accept "1:30", "00:01:30", "90", "90s", or a raw number. Clipped downloads of long videos are dramatically faster — a 30s slice of an hour-long video typically lands in ~15s vs ~60s for the full file. subtitles ignores the range and returns the full transcript.

Caching. Each unique (video, format, quality, range) tuple is cached for ~30 days in R2 + Redis. Repeat downloads within that window are instant. The returned URL is stable and safe to share or embed for the lifetime of the cache.

Limits. Capped at 3-hour source duration and 500 MB output. YouTube only — TikTok / Instagram aren’t supported here (their platforms reject the same infrastructure). Live, private, age-restricted, and region-blocked videos are rejected with specific error messages.

Hourly cap by plan. Per user (shared across all of your API keys): Starter 10/hr, Professional 100/hr, Ultimate unlimited. Cached repeats count; failed downloads don’t. When the cap is reached the tool returns an error widget with the wait time and a link to Subscription settings to upgrade.

The tool is most powerful when chained with other Algrow tools. Common patterns:

Outlier Autopsy

Use search_viral_videos with sort_by=outlier_score and min_outlier_score=3 to surface videos beating their channel's average by 3×+. Run analyze_video on each, asking "What does this do in the first 5 seconds, and how is it different from the channel's typical opener?". Algrow knows which videos overperformed; the analysis tells you why.

Niche Hook Pattern Extraction

Pair with search_viral_videos filtered by your niche and uploaded_within_days=7, top 10 by views. Run the same hook-focused analyze_video prompt across all 10. Claude synthesizes the cross-video patterns — cold-opens vs. question hooks vs. shock cuts, average length, on-screen text usage — into a single research memo.

Thumbnail-to-Hook Bridge

Use search_by_thumbnail to find videos with visually similar thumbnails to a draft or competitor cover. Then analyze_video on the top results with the prompt "How does the opening 3 seconds pay off the thumbnail promise?" — surfaces the gap between thumbnail bait and actual hook delivery.

Transcript + Vision Triangulation

Run get_youtube_video_data with include_comments=true, then analyze_video on the same URL. Claude reasons across visual analysis, public metadata, and audience reactions.

Competitor Folder Watch

For channels in a saved folder (get_folder), spot any with elevated views_24h, find the spiking upload via get_channel_videos, and analyze_video it with "What's the hook, format, and pacing — anything new vs. their previous style?". Pairs perfectly with a recurring weekly briefing.

Hook-to-Voiceover Pipeline

End-to-end content R&D in one chat: analyze_video on a top performer, ask Claude to draft 3 alternative hook scripts in the same style, then generate_tts on each so you can A/B-test voiceovers in your own video.