Video → LLM Analyzer
Pricing
from $0.008 / analysis completed
Drop a video, give it a prompt, and have an LLM analyze it. Results land in your Notion page so Claude Code can pick up where you left off. Uses Apify MCP Connectors to write to Notion on your behalf — no Notion API key required.
New here? Keep the example video and prompt and click Start to see how it works, or replace the example with your own video. Your analysis appears on the Output tab.
Easiest path: drop a file (MP4/MOV/WebM, ≤1 GB, ≤10 min — lift the 10-min cap with Process videos longer than 10 minutes in Advanced; costs more). Any video file works regardless of where it came from.
Or paste a URL from one of these platforms: YouTube + Shorts, TikTok, X (Twitter), Instagram.
Anything else — Vimeo, Dailymotion, Facebook, Reddit, Twitch, Bilibili, VK, and any other site not listed above — needs to be downloaded and uploaded as a file. We can't reliably download from these from the cloud yet. If you paste an unsupported URL, we'll tell you in about a second — before any compute runs, so you're not charged.
Plain English. "Summarize the main points", "List every product mentioned with timestamps", "Describe each scene and who's speaking." The more specific, the better the result.
For YouTube links: a visual prompt — one that mentions scenes, frames, what's on screen, who appears, etc. — automatically pulls real video frames (not just the thumbnail) as long as the clip is 10 minutes or shorter. For longer YouTube videos a visual prompt falls back to the transcript + thumbnail (the run says so in softErrors and the log); to force full-frame analysis on a long video, turn on Process videos longer than 10 minutes in Advanced (costs more) or download it and upload the file instead. Purely spoken-word prompts always use the faster, cheaper transcript path.
Which model family analyzes the video. No API keys needed — this Actor routes through Apify's hosted OpenRouter integration, so you only pay your Apify account at OpenRouter's actual cost (rounded to the nearest $0.00001). Pick the family here; for a specific model override, see the Model field below.
Pick Default to use the provider's standard model — Claude Sonnet 4.5, GPT-4o mini, or Gemini 2.5 Flash, depending on the provider above. Or pick a specific model. ⚠️ If your model is from a different family than the provider above (e.g. provider=Claude + model=openai/gpt-4o), the model's family wins and the run bills through that path. Keep the model family matching the provider, or leave on Default.
Only used when the provider above is Claude. Ignored for GPT and Gemini. Skip this unless you want to spend your own Anthropic credits instead of Apify's hosted OpenRouter (the default).
When filled AND the provider is Claude, the call hits Anthropic directly with your key — useful for Anthropic-specific features (prompt caching, extended thinking) or to bill your own Anthropic account instead of Apify's OpenRouter passthrough. The key lives in this Actor's memory for the run and is then discarded. Get a key at console.anthropic.com/settings/keys.
Skip this — most prompts don't need spoken words. Frames + the video's metadata cover ~90% of analysis prompts. Add a Whisper key (paste an OpenAI key from platform.openai.com/api-keys) only when the analysis genuinely depends on what people said. ~$0.006 per minute of audio.
This is the one piece that still needs a raw API key for now — no MCP server wraps Whisper yet.
Only used with a Whisper key. The language people speak in the video, as a two-letter code such as en, cs, de, or es (en-US also works; English doesn't). Leave empty and Whisper guesses from the first 30 seconds. Set it if a run's softErrors.whisper says the transcript repeats itself, or the transcript is in the wrong language (accented English is sometimes heard as another language).
Turn this on to get the analysis back without connecting anything — good for your first run, when you don't have a Notion / Slack / Sheets connection yet. It's pre-filled on in this form so a fresh run works with no setup; turn it off once you pick a destination below. Calling from an agent / API / Claude Code? The form pre-fill doesn't apply to API calls — pass skipDestination: true explicitly to get analysis-only, or pass a destination with skipDestination: false. Leaving the destination below empty without turning this on will stop the run and ask for a destination. When on, the destination + page/channel/sheet + write-behavior fields below are ignored and the analysis lands in this run's dataset row and OUTPUT record.
Optional. To run with no connection, turn on 'Just give me the analysis' above and leave this empty. Otherwise, pick a saved connection to auto-write results into Notion (fully tested), Slack, or Google Sheets (both functional, lighter-tested).
Don't have a connection? Click the field → Create new → log in to the app and authorize. Seeing Connectors tagged 'Not eligible'? Those are for other services (GitHub, Supabase, Sentry…) that don't write to Notion/Slack/Sheets — picking one is safe, we catch it before the LLM call and exit with no LLM cost incurred.
Where exactly within the connection. Format depends on what you picked:
| If you picked… | Paste this | Example |
|---|---|---|
| Notion | Page URL (not a database) | https://www.notion.so/My-Notes-1234abcd5678 |
| Slack | Channel name or ID | #video-notes or C012AB3CD |
| Google Sheets | Spreadsheet URL | https://docs.google.com/spreadsheets/d/<id>/edit |
Slack needs the Apify Slack app in the channel. Notion appends a section by default, or creates a new child page if your prompt asks for one. Sheets always appends a row.
Default (Smart) is right for nearly everyone. Reads your prompt — if you said "new page" / "separate section", creates a new container; otherwise appends to what's there. Pin one of the others if you want to override prompt-detection.
Leave at 0 for adaptive (~1 frame per 5s, between 4 and 12 frames). Override only if you have a specific reason — more frames = higher LLM cost. Note: very high frame counts can produce a payload too large for the LLM to accept; if you set this higher than 12 the Actor will cap it at 12 and say so in the output (softErrors.framesCapped).
YouTube tip: captioned YouTube videos are normally read via transcript + a single thumbnail (cheapest). Setting this above 0 — or writing a visual prompt (see the prompt field above) — makes the Actor download the video and extract real frames instead, as long as the clip is within the 10-minute limit (or Process videos longer than 10 minutes is on). On a captioned video this keeps the captions too, so you get frames and transcript.
⚠️ Off by default on purpose: the 10-minute cap keeps run costs predictable. Turn this on to process longer videos — the run will cost more, and the cost is harder to estimate up front: audio transcription bills per minute, downloads and compute take longer, and a 1-hour video costs roughly two to three times a 10-minute one. Strongly recommended: set Maximum cost for this run below — once the Actor reads the video's real length, it checks the worst-case cost against your cap and stops before transcription and the LLM call. Limits that still apply: 1 GB file size; transcription stops past ~69 minutes of audio (the run continues without it and says so in softErrors). Details in the README under Long videos.
By default, if you re-run with the exact same prompt + video + destination + behaviour as a previous run, the Actor replays the previous result instead of paying for the LLM call again. Set this to true to force a fresh analysis — useful if the source video has been re-uploaded under the same URL or you want a different angle on the same clip.
Set a ceiling on the LLM + transcription cost of this run. Once the Actor has downloaded and read the video (so it knows the real length and frame count), it computes a worst-case cost. If that estimate exceeds this cap, the run aborts before transcription and the LLM call — no LLM or Whisper cost incurred. This caps only the LLM/Whisper portion — Apify compute and the flat $0.015 per-run fee are billed on top (see the README pricing table). 0 means no cap — the explicit override if the worst-case estimate refuses a run you want anyway (the estimate is deliberately conservative). If you also set Maximum total charge in the Run dialog, the tighter of the two applies. As a guide, the LLM+Whisper portion is under $0.05 for most captioned YouTube videos and tops out around $0.15 for a 10-min upload on Claude Sonnet. Set $0.20 for headroom on typical videos; $0.50 for long uploads. If you turned on 'Process videos longer than 10 minutes', set this — long runs are exactly where costs get hard to predict (a 1-hour video with Whisper adds ~$0.36 of transcription alone).
