ytmusic-mcp
Enables searching YouTube Music and managing playlists: create/delete playlists, add/remove/reorder tracks, and more via natural language.
README
YouTube Music MCP Server (personal, single-user)
An MCP server wrapping ytmusicapi so Claude (and optionally ChatGPT) can
search YouTube Music and manage playlists: create/delete, add/remove tracks,
reorder tracks (top/bottom/before another track).
Scope of this setup: built for exactly one user (you), hosted on your own Ubuntu box, protected by an OAuth 2.1 login. Recommended to run against a secondary Google account, not your main one, since a leaked token means full playlist read/write access to whatever account is authenticated.
1. Install
pip install -r requirements.txt
2. Auth against YouTube Music (secondary account)
Log into your secondary Gmail/YouTube account in the browser you use for this, then run:
python3 setup_auth.py
Follow the printed steps (DevTools → Network tab → copy request headers
from a music.youtube.com request). This writes browser.json, which
server.py reads automatically. Cookie-based auth can go stale — re-run
this if tools start failing.
Paste the headers, then press Enter on one additional empty line. Firefox's structured request-header JSON export is also accepted. This project uses browser-cookie authentication only; OAuth credentials are rejected.
3. Local test (optional, before deploying)
python3 -c "from server import get_client; print(get_client().get_library_playlists(limit=5))"
4. Deploy on your Ubuntu server
4.1 Copy the project over
sudo mkdir -p /opt/ytmusic-mcp
sudo cp server.py oauth_provider.py setup_auth.py requirements.txt .env.example /opt/ytmusic-mcp/
sudo cp browser.json /opt/ytmusic-mcp/ # the file setup_auth.py made
cd /opt/ytmusic-mcp
sudo pip install -r requirements.txt --break-system-packages
4.2 Configure
sudo cp .env.example /opt/ytmusic-mcp/.env
openssl rand -base64 32 # generate the OAuth login password
sudo nano /opt/ytmusic-mcp/.env # set public URL, username, and password
Leave MCP_HOST=127.0.0.1 — the app must never bind directly to the public
interface. nginx is the only thing facing the internet.
Both direct Python runs and the systemd unit load this same .env file.
Values already exported by the shell or supplied by systemd take precedence.
nginx does not read dotenv files, so if you change MCP_PORT from 8787,
also change the port in proxy_pass in nginx-ytmusic-mcp.conf.
Set YTMUSIC_LOG_LEVEL=DEBUG for detailed application diagnostics. Dependency
protocol logs remain at MCP_LIBRARY_LOG_LEVEL=INFO by default so journald does
not print raw SSE/JSON-RPC payloads. Tool failures are logged with tracebacks and
are also returned to the MCP client as tool errors.
4.3 Run as a systemd service
sudo cp ytmusic-mcp.service /etc/systemd/system/
sudo nano /etc/systemd/system/ytmusic-mcp.service # set User= to your linux user
sudo systemctl daemon-reload
sudo systemctl enable --now ytmusic-mcp
sudo systemctl status ytmusic-mcp
journalctl -u ytmusic-mcp -f # tail logs
4.4 nginx — HTTPS MCP and OAuth routes
sudo apt install nginx certbot python3-certbot-nginx
sudo cp nginx-ytmusic-mcp.conf /etc/nginx/sites-available/ytmusic-mcp
sudo nano /etc/nginx/sites-available/ytmusic-mcp # replace mcp.yourdomain.com
sudo ln -s /etc/nginx/sites-available/ytmusic-mcp /etc/nginx/sites-enabled/
sudo nginx -t && sudo systemctl reload nginx
sudo certbot --nginx -d mcp.yourdomain.com # gets the TLS cert, edits the config
You need a domain/subdomain pointed at your server's public IP (any DNS A record works), and ports 80+443 open in your firewall. Port 80 exists only to redirect to HTTPS and serve certbot's renewal challenge — nothing else is served there.
This config exposes /ytmusic-mcp and the small set of OAuth discovery,
registration, authorization, token, and login routes it requires. Other paths
(/, /admin, /.env, etc.) return a bare 404 before hitting the app.
It rate-limits MCP and login attempts (30/min, burst 10).
4.5 fail2ban — auto-ban repeat offenders
sudo apt install fail2ban
sudo cp fail2ban-filter-ytmusic-mcp.conf /etc/fail2ban/filter.d/ytmusic-mcp.conf
sudo cp fail2ban-jail-ytmusic-mcp.conf /etc/fail2ban/jail.d/ytmusic-mcp.conf
sudo systemctl restart fail2ban
sudo fail2ban-client status ytmusic-mcp # confirm the jail is active
Any IP racking up 5 unauthorized/probing requests in 10 minutes gets banned for 24h.
4.6 Verify from outside
curl -i https://mcp.yourdomain.com/ytmusic-mcp # expect 401 with OAuth metadata
curl -i https://mcp.yourdomain.com/anything # expect 404 (nothing else exposed)
Your MCP endpoint: https://mcp.yourdomain.com/ytmusic-mcp
5. Connect from Claude.ai
- Settings → Connectors → Add custom connector.
- URL:
https://mcp.yourdomain.com/ytmusic-mcp - Normally leave Client ID and Client Secret empty; Claude will use OAuth
Dynamic Client Registration. To use the advanced fields instead, configure
the matching
OAUTH_CLIENT_ID,OAUTH_CLIENT_SECRET, and Claude callback URI inOAUTH_REDIRECT_URISfirst. - Save, complete the private username/password login, then enable it in a conversation via + → Connectors.
6. Connect from ChatGPT
ChatGPT's remote-connector/custom-header support has been shifting between
plan tiers and connector types — worth checking OpenAI's current docs for
the exact click path when you get there. The server side doesn't change:
it's a standard endpoint at https://mcp.yourdomain.com/ytmusic-mcp using
the standard MCP OAuth discovery and authorization flow.
Security summary
- The OAuth login protects access to your YouTube account (via
browser.json) — use a generated password and keep it private. - App binds
127.0.0.1only; nginx is the sole public-facing surface and exposes only the MCP resource and required OAuth routes on:443. - Rate limiting + fail2ban blunt scanning/brute-force attempts.
- Using a secondary Google account caps the blast radius of a leak to that account's playlists, not your main one.
- Rotate the OAuth password any time you suspect exposure: regenerate, update
.env,sudo systemctl restart ytmusic-mcp. browser.jsoncan expire — re-runsetup_auth.py(on your dev machine, then re-copy to the server) if calls start failing with auth errors.
Real-account integration tests
The integration suite uses browser.json and performs real YouTube Music
writes. It creates a uniquely named private playlist and deletes it during
cleanup, including when a test fails. Run it explicitly with:
python3 -m pip install -r requirements-test.txt
pytest --run-youtube-integration -v
Without --run-youtube-integration, these account-mutating tests are skipped.
Tools exposed
| Tool | What it does |
|---|---|
search_music |
Search songs/videos/albums/artists/playlists |
list_my_playlists |
List your playlists |
get_playlist |
Get a playlist's tracks (includes setVideoId, needed for remove/reorder) |
create_playlist |
Create a new playlist |
delete_playlist |
Delete a playlist |
rename_playlist |
Edit title/description/privacy |
add_tracks_to_playlist |
Add tracks by videoId, optionally to the top |
remove_tracks_from_playlist |
Remove tracks by setVideoId |
reorder_playlist_track |
Move a track before another track, or to the end |
move_track_to_top |
Move a track to position 1 |
move_track_to_bottom |
Move a track to the last position |
ID note: videoId identifies a song (used to add); setVideoId
identifies one occurrence of a track inside a playlist (used to remove
or reorder, since the same song can appear twice with different
setVideoIds). You never need to look these up manually — ask by song name
and the assistant calls get_playlist first to resolve the right ID.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。