tubeai-app/tubeai-skills/tubeai-mcp

Claude Code and agent skill by tubeai-app/tubeai-skills

Trust tier 3
  • community
Source check pending

Checking that the SKILL.md file is still in the repository.

YouTube research through the TubeAI connector (MCP). Explores 400M+ YouTube videos and 4M+ channels to find video ideas and outliers, research niches and competitors, work out titles, study transcripts and draft scripts, makes thumbnails, and saves ideas and folders to the user's TubeAI workspace. Helps the user connect and sign in when they want it. Use when the user wants video ideas, what's working in a niche or on a channel, competitor research, title or thumbnail help, a transcript or a script draft, or mentions TubeAI.

Install

npx vskill@latest install tubeai-app/tubeai-skills/tubeai-mcp

SKILL.md

TubeAI connector

TubeAI connects to Claude as a connector (an MCP server) at https://beta.tubeai.app/api/mcp. It gives Claude a database of 400M+ YouTube videos and 4M+ channels, plus the user's own TubeAI workspace: ideas, folders, boards and past work. Use it to explore YouTube data, find video ideas and draft scripts with the user.

Assume the user isn't technical: do the technical parts yourself, give click-level steps when they have to act, and talk in plain language.

Once connected, TubeAI sends its own instructions and a full description of every tool; follow those for each tool's exact inputs. This skill covers getting connected, which tools to use for each job, and the rules that keep the user's account safe.

Connect first

TubeAI is optional, but this skill runs on it. Before the first TubeAI task in a session:

  1. Check the connection with a workspace_account call. When it answers with the user's plan, TubeAI is connected.
  2. If it isn't, offer it in a line: what it gives them (video ideas, packaging and what's working in their niche) and that it needs a free TubeAI account. If they want it, walk them through one of the two ways:
  • In the Claude app (claude.ai, the desktop app or mobile). A connector added there also works in Claude Code when they're signed in with the same Claude account.
  1. Open Settings → Connectors (under Customize).
  2. Click Add → Add custom connector, name it TubeAI, and paste https://beta.tubeai.app/api/mcp as the server URL. Leave Advanced settings empty.
  3. Click Add, then Connect. TubeAI's authorization page opens: check the account shown and press Approve. They need to be logged in to TubeAI in that browser first.
  • Or in the terminal (Claude Code): run claude mcp add --transport http --scope user tubeai https://beta.tubeai.app/api/mcp for them. Then they type /mcp, pick tubeai, choose to authenticate, and sign in to TubeAI in the browser that opens.
  1. Never add a second TubeAI connection. If TubeAI is already added (in the app's Connectors, or in /mcp in Claude Code), it only needs signing in. On Claude Team and Enterprise plans only admins can add connectors in the app, so use the terminal line there.
  2. Tools not showing up: in the Claude app, open the TubeAI connector in Settings → Connectors and choose ⋮ → Refresh tools list, or Disconnect and Connect again. A connector added during a session may only show up in a new chat.
  3. Check again with workspace_account. If they'd rather not connect, carry on without it (WebFetch, Claude in Chrome and yt-dlp for research), and don't offer it again unless they ask.

Budget and credits

  • Every tool call counts against the plan's daily request budget, which the TubeAI app's database pages share. It resets at 00:00 UTC. Going over it locks the user's database access, in the app too, for 2 days. So:
  • make one well-aimed call rather than many small ones
  • keep limit small, and ask for extra fields only when you need them (results are slim by default)
  • for many channels at once, use one explore_videos call with a channel set and perChannel, never one call per channel
  • get the niche labels from explore_niches once, then reuse them
  • workspace_account shows the plan, the credits, what each paid tool costs, which features the plan lacks, and today's budget. Check it before anything that spends credits, when a tool says a feature is locked, or when the user asks about their plan.
  • Spending credits needs the user's explicit OK after they've seen the cost: thumbnail_generate, thumbnail_edit and a board's dataSources (Google Trends + X posts). Everything else is free, the other thumbnail tools included.
  • A section or list that comes back locked names the plan that includes it. Mention the upgrade only when it matters to what the user wants.

Playbooks

Find video ideas (for the user's channel, a niche, or a saved folder)

  1. mini_ideas with exactly one scope: channel (an @handle or URL), niche (a label from explore_niches) or folderId (from workspace_folder_get). Defaults: the last 90 days, videos at 2× their channel's average or more, long-form only (includeShorts: true adds Shorts).
  2. It returns the outlier videos, the title formats that perform best in that niche, and the words those outlier titles share. Draft ideas from these patterns: a title, the format it borrows, and the angle that makes it the user's own.
  3. Check the draft titles with mini_titles (titles, up to 20). It marks which words and phrases carry proven performance.
  4. Save the keepers (see Save the keepers).

Research a niche

  1. explore_niches without arguments lists every niche with its outlier stats. With niche, it gives that niche's top channels, top videos and sub-niches.
  2. explore_videos with niches and a minDate window, sorted by outlier_multiplier, shows what's working there now. minDuration: 181 leaves Shorts out.
  3. explore_channels with niches plus size and performance filters (minSubscribers, minAvgViews, minOutlierRate as a percent, isActive) finds the channels worth watching.

Study a competitor, or the user's own channel

  • explore_channel gives the profile, plus the include sections: recent and outliers by default, and momentum and performance when the plan has them.
  • explore_similar_channels finds channels like it.
  • A creator's recent outliers: explore_videos with channelHandles: ["@x"], minOutlier: 1, minDate six months back, sortBy: "outlier_multiplier".
  • Many channels at once: one explore_videos call with channelHandles or channelIds (up to 50) and perChannel (the top N videos of each).

What's taking off right now

  • explore_videos with similarToHandle: "@x" (the 100 channels most like it), minOutlier: 2, minDate 30 days back, sortBy: "published_at".

Titles and packaging

  • mini_titles with show: ["formats", "words"] gives proven title structures and high-performing title words. Narrow them with search and (formats only) niches, and rank them best_performing, most_used or trending.
  • To see the packaging: showThumbnails: true on explore_videos (the first 12 videos), or showThumbnail: true on explore_video.
  • explore_similar_thumbnails finds videos whose thumbnails look like one video's thumbnail. It only works for videos TubeAI has analyzed.

Thumbnails (they cost credits, so confirm the cost first)

  1. Agree on the concept with the user: what the thumbnail shows, and any text on it.
  2. thumbnail_generate with the video's title or idea (videoIdea) and the concept (thumbnailConcept). count is 2 or 4, and only one generation runs at a time, so ask for every option in one call. With channelHandle, TubeAI matches that channel's style from its real thumbnails: then describe only the idea, never colors, fonts or layout.
  3. thumbnail_status, called once, shows the gallery, which fills in by itself in 1–2 minutes. thumbnail_view lets you see the images and gives their historyIds.
  4. thumbnail_edit changes one. The user names it by its number in the gallery, which is the historyId at that position.
  5. The user's own face: thumbnail_face_upload shows an uploader. When they say it's done, generate with useUploadedFace: true.

Style matching and the user's face depend on the plan. A locked feature says which plan includes it.

Draft a script

  • Start from the idea the user picked and the patterns behind it: the hook, the promise the title makes, and the beats that deliver it. Draft it with the user one part at a time, in their channel's voice.
  • To study how a hit is scripted, or how a creator actually talks, get transcripts with mini_script: videos (1–5 ids or URLs, any public video) or channel with count (its latest uploads). Transcripts TubeAI already has are free; new ones count against a daily transcript cap. An entry marked aiSummary: true is a summary of the video, not the words spoken, so never learn a voice from it.
  • The user's own past work in TubeAI helps: workspace_history with section: "scripts" (or viral, metadata, analytics) lists what they made in the app. A script's historyId is what workspace_idea_save takes as scriptProposal.

Save the keepers

  • Ideas: workspace_idea_save creates a private draft in the user's My Ideas planner. Never say it's published: the user reviews and publishes it in TubeAI.
  • A title of 24–58 characters is the planner's sweet spot.
  • Add 2–4 sentences of notes on the angle and why it works now.
  • Add up to 6 reference videos, only ones your research returned.
  • ideaId edits an existing idea, and each list you pass replaces the stored one. workspace_idea_get lists or opens ideas.
  • Channels and videos: confirm the folder name and its contents with the user, then use workspace_folder_create or workspace_folder_add. A folder holds channels or videos, never both, and takes up to 200 items per call. workspace_folder_get lists folders and their ids.
  • Anything that didn't save comes back under unparseable, wrongType or notFound. Tell the user instead of reporting a clean save.
  • Boards, when the plan has them: workspace_board_create builds a planning board from a folder, a list of videos or the channel's niche, and labels it for free. A niche scan can come back as status: "scanning". Tell the user it's in progress, then call again with the returned resumeOperationId and every field of resumeWith.

Reading the numbers

  • outlier_multiplier: a video's views against its own channel's average (3 = three times the usual).
  • runner_multiplier: views against the channel's subscriber count.
  • outlier_rate (channels): the percent of a channel's videos that are outliers.
  • Shorts run up to 3 minutes, so minDuration: 181 leaves them out.
  • unknownNiches and unresolvedHandles list what didn't match. Say so rather than treating the results as complete.

Good to know

  • The tool list can show TubeAI's tools with a prefix (like mcp__tubeai__explore_videos). Use the name it shows.
  • explore_videos refuses a call without at least one narrowing filter: search, niches, a channel set (channelIds, channelHandles or similarToHandle) or minDate.
  • Name channels by @handle, channel URL or UC id, and videos by id or URL.
  • Results page with cursor and nextCursor.
  • A query that comes back as too broad: narrow it (a shorter window, fewer channels, a smaller limit) and try once more.
  • A tool says it needs authentication, or the TubeAI tools are missing: go back to Connect first.
  • The daily budget is used up: it resets at 00:00 UTC. Carry on with the fallbacks and tell the user in a line.
  • A tool says database access is paused: the budget was exceeded. Tell the user; TubeAI support can lift it early.

Security

  • Security scanPassed, 95 of 100Pattern checks on the skill's files.
  • Model reviewPassed

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

Repository stars
60
Repository forks
15
7-day trend score
0

Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/tubeai-app/tubeai-skills/tubeai-mcp/badge)](https://verified-skill.com/skills/tubeai-app/tubeai-skills/tubeai-mcp)

Something wrong with this skill or its page? Tell us.