MCP · Agent Skills · Claude Code
YouTube Skills vs MCP: Give Claude Code, OpenClaw or Hermes Agent Transcripts and Channel Research
To let an agent read YouTube transcripts and research a channel, give it an MCP server (tools) or an Agent Skill (instructions that call an API). We ran both in Claude Code on the same channel and compare what each did.
· 7 min read · YTAPI
You can give an agent YouTube access two ways. An MCP server adds tools the agent can call (get_transcript, get_channel_videos). An Agent Skill is a folder with a SKILL.md that tells the agent how to do a job, usually by running curl against an API. In Claude Code, OpenClaw and Hermes Agent both work. MCP is the shorter setup. A skill carries a method, such as how to decide which uploads broke out.
We make YTAPI, a transcript API, so both setups below use it. We tested them in Claude Code 2.1.325 on 2026-11-15. We did not install OpenClaw or Hermes Agent, so their sections come from their official docs and are marked as untested.
Skill or MCP: what differs
| MCP server | Agent Skill | |
|---|---|---|
| What it is | Tools with typed inputs, served by a remote or local server | Markdown instructions, optionally with scripts |
| What the agent needs | An MCP client | A shell (the YTAPI skills call curl) and an environment variable for the key |
| Sign-in | OAuth in the client, or a key in a header | A key in YTAPI_API_KEY |
| Adds to context | Tool definitions for the server | A short description, with the full text loaded when used (Hermes docs call this progressive disclosure) |
| Carries a method | No. The agent decides what to call | Yes. Steps, thresholds, report format |
They combine. The ytapi-youtube-channel-analysis skill says to use the YTAPI tools when the session already has them, and fall back to curl when it does not (SKILL.md).
Claude Code
MCP
The interactive route is the plugin, with sign-in through /mcp; YouTube Transcripts in Claude walks through it. For a scripted setup, the server also accepts a key in a header. We added it to a project's .mcp.json with claude mcp add:
claude mcp add --scope project --transport http ytapi https://api.ytapi.dev/mcp \
--header 'Authorization: Bearer ${YTAPI_API_KEY}'Claude Code asks you to approve a project-scoped server the first time. In a headless run, claude -p with --mcp-config connected straight away, and the server exposed 10 tools: get_balance, get_channel, get_channel_videos, get_comments, get_playlist_videos, get_transcript, get_video_details, get_video_heatmap, get_video_info, search_youtube.
Skill
The skill is in the public ytapi/youtube-skills repo. The installer from skills.sh copied one skill into the project:
npx skills add ytapi/youtube-skills --skill ytapi-youtube-channel-analysisIt placed SKILL.md and a references/auth-setup.md under .claude/skills/ytapi-youtube-channel-analysis/. Before installing, the tool printed third-party risk ratings for the repo (one scanner showed "Safe", another "1 alert", a third "Med"). That fits what the skill does: it runs shell commands with your API key, so read it before you install it. Set YTAPI_API_KEY in the environment and the agent uses it. If it is unset, the skill's setup file has the agent offer to create a free account from your email.
The same task, both ways
We asked each setup to research @3blue1brown (8.76M subscribers, 247 videos at test time): find the biggest breakout, read its transcript, and describe the first minute.
MCP, no skill
The agent made four tool calls: get_channel, get_channel_videos (popular), get_channel_videos (newest) and get_transcript with max_chars: 4000. It took 23 seconds. It judged breakouts against the median of the 30 newest uploads, about 1.7M views, and picked "Large Language Models explained briefly" (8.5M views, 7:58 long, about 5x that median). Its summary of the opening:
Opens with a thought experiment, not a definition. You find a movie script where the AI's reply is torn off, and you have a machine that predicts the next word. [...] It says that is what a chatbot does by 0:33. [...] There is no intro, no channel branding and no outline of what's coming.
We checked this against the transcript: the first caption starts "Imagine you happen across a short movie script...", and the chatbot link comes in the caption at 0:33. The agent also said it had read only about 4,000 of 7,800 characters, which is the cap it set itself.
Skill, no MCP
We ran the skill with the MCP server turned off, so the agent used curl. It took 59 seconds and, by its own count, 6 requests: profile, newest page, popular page, video details, two transcripts. It compared videos with the uploads published around the same time instead of the newest ones. Its table, trimmed:
| Video | Views | Length | Multiple of nearby median |
|---|---|---|---|
| Transformers, the tech behind LLMs | 11.67M | 27:14 | about 4.5x |
| Large Language Models explained briefly | 8.5M | 7:58 | about 3x |
| How are holograms possible? | 7.1M | 46:24 | about 2.7x |
For the top video it also reported exact counts from the video details (about 12.2K views per day, 239.5K likes, a 2.05% like rate) and noted that the popular page is mostly years-old evergreen videos, which show what lasted and not what is breaking out now.
What was different
The two runs named different videos as the biggest breakout. The MCP-only agent made up its own baseline from the newest uploads, so a 2024 video that gathered views for two years was not compared fairly with recent ones. The skill told the agent to compare each video with uploads from around its own date and to use exact counts for the finalists. That is a method, and an MCP server does not contain one.
The MCP run was also simpler to start: no shell, no environment variable, and a prompt of one paragraph. Both runs gave plausible answers. We did not rerun them to measure how often each picks the same video, so treat the comparison as one run each.
For a longer channel study, with growth stories and a reusable playbook, see How to Study a Competitor's YouTube Channel with Claude. For other MCP servers that do transcripts, see Best YouTube MCP Servers compared.
OpenClaw (not tested)
We did not install OpenClaw. Its install page (read 2026-11-15) lists Node 24.16 or newer and an onboarding step after the installer, and this machine runs Node 22.
- Skill. The YTAPI repo's README gives
npx clawhub@latest install ytapi. The OpenClaw skills docs also listopenclaw skills install skills-sh:owner/repo/slugfor skills.sh repos. - MCP. The MCP registry docs describe
openclaw mcp addandopenclaw mcp set <name> <json>, and the transports page shows a remote server as aurl,transport: "streamable-http"andheaders. That fits YTAPI's endpoint with a bearer key, but we have not run it.
Hermes Agent (not tested)
We did not install Hermes Agent either. The install guide uses a curl | bash script that also sets up Python, Node and a browser.
- Skill. The YTAPI README gives
hermes skills install skills-sh/ytapi/youtube-skills/skills/ytapi. The skills docs show the sameskills-sh/form and say skills live in~/.hermes/skills/. - MCP. The MCP docs show a remote server under
mcp_serversin the config file withurlandheaders, so the same bearer-key setup applies.
If you try either, check the exact flags in the linked docs first; both projects change often.
Which to pick
- You use Claude Code and want it working in two minutes: the plugin or the MCP server.
- You want a repeatable method, such as breakout analysis with a report format: add a skill. It works with MCP tools or without them.
- Your agent has no MCP support but has a shell: the skill is the only route.
- A server or cloud agent: both go through the YTAPI API, so neither depends on fetching YouTube from that machine's IP.
Cost and limits
- Requests cost one credit each. Video basic info (
get_video_info) and failed requests are free. By the request lists above, the MCP run used about 4 credits and the skill run about 6. - New accounts get 200 free credits. Until you buy a pack, a key is limited to 1 request per second and 100 per day. Packs start at $9 for 2,000 credits. See pricing.
- The 25-video cap applies to one-click exports in the dashboard and Chrome extension, not to the API or MCP.
- Videos without captions, private, members-only and age-restricted videos return an error. Channel listings give rounded view counts, and the pages hold regular uploads only, not Shorts or live streams.
- We ran these tests with an API key in a header, not the OAuth sign-in, because they ran headless.
FAQ
Do I need both a skill and MCP?
No. Either works alone. The skill adds a method, and the MCP server adds tools, and the YTAPI skills use the tools when they are present.
Is a skill safe to install?
It can run shell commands with your key in the environment, so read the SKILL.md and any scripts first. The YTAPI skills only call api.ytapi.dev with curl; you can check that in the repo.
Does the skill work without a key?
It needs one. If YTAPI_API_KEY is unset, the skill has the agent offer to create a free account and key from your email. We did not test that flow here.