Add a video to a playlist by playlist id or exact title. Optionally at a position (0 = first).

Which AI providers are connected (Claude, OpenAI, Azure OpenAI, Gemini, OpenRouter, Groq, Ollama, custom) and which model each job uses: text (writing, analysis), vision (judging thumbnails), image (making thumbnails), transcribe. Keys are never shown.

Full performance picture for one video: views vs this channel's usual (outlier score), watch time, average view percentage, hook retention at 30 s and the retention curve, traffic sources, YouTube search terms, subscribers gained, SEO score. With an AI model connected it adds a diagnosis: what's working, what's hurting, title variants and thumbnail ideas ready to A/B test, and next actions.

Show whether Google OAuth is configured, which YouTube channels are connected (and whether their sign-in is healthy), the default channel, today’s quota use, and known project limitations. Call this first in a new conversation.

Compares first-week views of the channel's recent public uploads by weekday and time of day they were published (in your time zone), relative to the channel median. Uses YouTube Analytics; needs 6+ videos older than a week.

Stop an upload job that has not finished. It can be resumed later with resume_upload.

Start Google sign-in to connect (or reconnect) a YouTube channel. Opens the browser on this computer and returns the sign-in link; the user picks the Google account and channel (brand accounts work). Waits up to wait_seconds for them to finish.

Next video ideas grounded in this channel's data: its outliers (what over-performed vs its usual), underperformers, and optionally a topic to research. With an AI model it returns ideas with angle, why now, thumbnail concept and confidence; otherwise the data for you to reason from.

Tests 1-3 challenger versions of a live video against its current version (A). The server rotates which version is live one full day at a time (Pacific days, matching YouTube Analytics), in balanced random order so weekdays even out, then compares daily views, watch time, average view percentage, subscribers or engagement, adjusting for the video's natural decline in daily views, with a significance test. At the end it restores A, waits ~3 days for analytics to land, and (if auto_apply_winner) puts a significant winner live. YouTube Studio's own Test & Compare (with impressions click-through rate) has no API; this works on any channel through the official APIs. Each variant can change title, thumbnail and/or description. Changes are public, so the user confirms first.

Create a new playlist on the channel.

Cuts start-end (seconds) from a local video into a vertical 1080x1920 MP4 for YouTube Shorts (max 180 s). layout crop fills the frame from the centre (talking heads); blur shows the whole frame over a blurred fill; original keeps the aspect ratio. Needs ffmpeg.

Scores 1-4 thumbnail images as a phone viewer sees them (readability at small size, emotion, subject clarity, fit with the title), ranks them and lists fixes. Needs a vision model (Claude, OpenAI, Gemini, Azure).

Delete an uploaded caption track (see list_captions for ids). Needs the user’s confirmation.

Delete a comment or reply that the channel itself wrote (other people’s comments are removed with moderate_comments). Needs the user’s confirmation.

Delete a playlist (the videos in it are not deleted). Needs the user’s confirmation.

Delete a saved template file.

Permanently delete a video from YouTube. Cannot be undone. Always needs the user’s confirmation.

Remove a connected channel and revoke this app’s Google access to it. Needs the user’s confirmation.

Download a caption track of your own video as SRT (prefers uploaded tracks, falls back to auto-generated). Costs 250 quota units.

Saves evenly spaced stills (or stills at given seconds) from a local video as 1280-wide JPGs, as thumbnail candidates or as references for generate_thumbnails. Needs ffmpeg.

Ranks the channel's videos by outlier score (views ÷ the median of neighbouring uploads of the same format). Shows what over-performed (repeat it) and what under-performed (fix packaging or drop the topic). Shorts and long videos are compared separately.

Gets the transcript with timestamps (from text you pass, a .srt/.vtt/.txt file, an uploaded video’s captions, or by transcribing a local file). When an AI writing model is configured (Azure OpenAI, OpenAI…), it returns a finished draft: 3 titles, description with chapters, tags, hashtags, pinned comment and thumbnail text, already merged with the template and checked against YouTube’s limits. Otherwise (or with draft: false) it returns a brief for you to write from. Show the draft to the user before using it.

Designs thumbnail concepts for a title (text overlay + image prompt) with the text model, then renders them with the image model (OpenAI gpt-image, Azure OpenAI image deployments, Gemini image). Optionally builds on a video frame. Saves 1280x720 JPGs under 2 MB, ready for set_thumbnail or create_ab_test. Without an image model it returns the concepts only.

Schedule, live variant and results so far (per-variant daily averages, lift vs the original, p-values, verdict). Results become reliable once every version has 3+ clean days; analytics lag about 2 days.

YouTube Analytics for the channel or one video. Reports: overview (views, watch time, average view duration, subscribers gained/lost, likes, comments, shares), daily (trend by day), top_videos (best performers with titles), traffic_sources, search_terms (YouTube search queries), geography, devices, demographics (age/gender), retention (audience retention curve, needs video_id), subscribed_status, or custom (your own metrics/dimensions). Defaults to the last 28 days. Data lags about 2 days.

The channel’s title, handle, description, keywords, country, default language, trailer, subscriber/view/video counts, banner and settings.

Status of an upload job (progress, errors, resulting video) and/or a video’s YouTube processing state (processing, failed, rejected), plus a check for the "locked to private" restriction. Can wait for changes.

Full details of one video: metadata, privacy and schedule, processing state, duration, captions flag and statistics.

A/B tests with their status (scheduled, running, collecting results, completed, stopped), live variant and winner.

Caption tracks on a video (language, name, auto-generated or uploaded, draft).

YouTube video categories you can assign (id and name) for a region.

List the YouTube channels connected to this server.

Comment threads on one video, or across the whole channel (newest first). Filter by moderation state (published, held_for_review, likely_spam) or search text. Includes replies.

List videos on the channel, newest first (including private, unlisted and scheduled ones), with privacy, schedule, duration and view counts. Paginate with page_token.

Videos in a playlist, in order.

All playlists on the channel with id, title, privacy and video count.

Metadata templates (title format, description footer, default tags, hashtags, category, playlists, voice guidance) available for uploads.

Recent upload jobs from this computer, newest first, with progress and results.

Publish held comments, hold comments for review, or remove (reject) them; optionally ban the author from commenting on the channel. Removing or banning needs the user’s confirmation.

Move a video to a new position in a playlist (0 = first). Works for playlists sorted manually.

Show exactly what would be sent to YouTube (title, description, tags, playlists) after applying a template, check it against YouTube’s limits and chapter rules, and score its SEO/packaging (0-100 with fixes). No API call, no quota.

Estimated YouTube API quota used today (units and uploads) for this Google Cloud project, and when it resets (midnight Pacific Time).

Remove a video from a playlist (the video itself is not deleted).

Post a public reply to a comment (comment_id = the top-level comment or thread id) or a new top-level comment on a video (video_id). Posts publicly as the channel, so the user confirms the exact text first.

Searches YouTube for recent videos on a topic and ranks them by breakout ratio (views ÷ the uploader's subscribers: above 1 means it reached far beyond its audience), with title patterns (length, numbers, questions, common words) and typical duration. Costs about 102+ quota units per call.

Resume a stopped, failed or cancelled upload from the last byte YouTube confirmed (for example after a quota reset or reconnecting the channel). Network drops and quota resets are also retried automatically.

Create or replace a metadata template, saved as JSON in the templates folder. Use {title} in title_format. extends inherits another template; tags and hashtags combine.

Scores a video's (or draft's) title, description, tags, hashtags and chapters against YouTube best practice, 0-100 with fixes in order of impact. No AI needed; 1 quota unit for a video_id, none for a draft.

Pick which connected provider and model handle each job, as "provider:model" (Azure: "azure:"), e.g. text "anthropic:claude-opus-5-5", image "openai:gpt-image-2", transcribe "groq:whisper-large-v3". Pass null to go back to the automatic choice.

Upload a channel banner (JPG/PNG, 16:9, at least 2048x1152, ideally 2560x1440, max 6 MB). Public, so the user confirms first.

Choose which connected channel is used when a tool call does not name one.

Upload a JPG or PNG (max 2 MB, ideally 1280x720) from this computer as the video’s thumbnail. The channel must be phone-verified.

Add or update the title and description shown to viewers in other languages, e.g. {"hi": {"title": "...", "description": "..."}}. Merged with existing translations. The video needs a default_language set (update_video).

Change who can see a video: private, unlisted or public, or schedule it with publish_at. Anything other than private needs the user’s confirmation.

Set the subscribe watermark shown on all channel videos (square JPG/PNG, at least 150x150, max 1 MB), or remove it with remove: true.

Stops a test and puts the chosen version live (A = original). Public change, so the user confirms first.

Finds the moments in a long video that work as standalone Shorts (hook in the first seconds, one idea, a payoff) with start/end times, hook line, title and a score. Use a local file (transcribed), an uploaded video's captions, or a transcript file. Then cut them with create_short.

Transcribe a video file on this computer with any OpenAI-compatible speech-to-text API (needs ffmpeg and a configured key). Saves an .srt you can pass to upload_video or upload_captions, and returns a timestamped transcript for writing metadata and chapters.

Drafts the video's title and description in other languages so it can reach viewers who search in them. Returns drafts only; apply them with set_video_localizations after the user approves.

Reads recent comments (one video or the whole channel) and returns sentiment, themes, unanswered questions with draft replies, comments worth a reply, spam or abuse to moderate, and video ideas the audience is asking for. Drafts only: nothing is posted. Without an AI model it returns the comments for you to triage.

Change the channel description, keywords, country, default language or the trailer shown to non-subscribers. Only the fields you pass change. These are public, so the user confirms first. (The channel name and handle can only be changed in YouTube Studio.)

Rename a playlist, change its description or privacy. Making it public or unlisted needs the user’s confirmation.

Change a video’s title, description, tags, category, languages, audience and other settings. Only the fields you pass change. Optionally apply a template to the fields you pass (title format; footer and hashtags when you pass description; template tags when you pass tags). Changing visibility to public/unlisted or scheduling needs the user’s confirmation (or use set_visibility).

Add a captions track (.srt, .vtt or other YouTube-supported format) to a video, from a file on this computer or inline content.

Upload a video file from this computer (file_path) or a public URL to YouTube with a resumable, checkpointed upload that survives network drops and restarts. Optionally schedule it (publish_at), set a thumbnail, add it to playlists and attach captions in the same call. Runs in the background: returns a job_id right away (or the finished video if it completes within wait_seconds). Poll get_upload_status. Private uploads start immediately; public, unlisted and scheduled uploads need the user’s confirmation first. Tip: write metadata first (generate_metadata, then preview_metadata).