ytdl-rmcp

yt-dlp search, download, metadata, delivery, and Plex workflows over MCP and CLI.

Written in Rust on the rmcp crate. yt-dlp and ffmpeg are auto-downloaded into a per-user cache on first run, so the host needs neither pre-installed — the one binary is the whole install.

30-second path: npx -y @dinglebear/rytdl setup -> configure a target path -> call youtube_search or youtube_probe; use youtube_download only after the destination and trust boundary are clear.

Status: production personal-media MCP server. Read-only search/probe/stats paths are safe; download, playlist, queue-drain, and tag-writing paths create or move state and are intended for trusted callers.

Not for: a generic web-downloader SaaS, a multi-tenant media ingestion boundary, a replacement for yt-dlp's upstream site handling, or an arbitrary filesystem writer for untrusted MCP callers.

Contents

• Naming
• Capabilities And Boundaries
• Install
• Quickstart
• Client Configuration
• Runtime Surfaces
• MCP Tool Reference
• CLI Reference
• Configuration
• Authentication
• Safety And Trust Model
• Architecture
• Distribution Contract
• Development
• Verification
• Deployment
• Troubleshooting
• Related Servers
• Documentation
• License

Naming

The crate and npm package use the *-rmcp family naming pattern, while the repository and runtime binary are rytdl so local shells get a short Rust-native command.

Capabilities And Boundaries

• Searches YouTube through yt-dlp without downloading media.
• Downloads audio, video, or both into a staging tree, tags audio metadata, and transfers the result to local, SSH, or rclone destinations.
• Optionally fingerprints audio through AcoustID/MusicBrainz and syncs completed audio downloads into a Plex playlist.
• Builds Plex playlists from successful transfer history and drains server-made retained-staging transfer manifests.
• Exposes an MCP App search UI for hosts that can render embedded widgets.
• Keeps a JSONL ledger for repeat-safe downloads and stats.

Features

• Audio, video, or both — audio-first by default, with separate targets for audio and video.
• Proper tagging — embeds title / artist / album / date and cover art, and organizes output as Artist/Title [id].ext so media servers (Plex, etc.) index it cleanly. A non-greedy Artist - Title parse recovers the artist from free-form video titles. Source .info.json, thumbnail, and description sidecars are preserved next to the media for future retagging/indexing. Common YouTube title noise like (Official Video), [Official Audio], and trailing channel handles is stripped from embedded title metadata by default.
• Self-contained paths — the binary downloads/caches yt-dlp + ffmpeg when run directly; the container image bakes in ffmpeg, fpcalc, SSH, and rsync for media-host batch jobs.
• Self-installing — ytdl-rmcp setup registers the server into Claude Code, Codex, and/or Gemini CLI via each tool's own mcp add.
• Robust transfers — local paths (/path) are copied in-process by the binary itself, SSH targets (host:/path) use rsync -a --partial --protect-args with an scp fallback when rsync is missing, and rclone targets (remote:path or rclone:remote:/path) use rclone copy. On transfer failure the local staging copy is kept for retry and recorded as a drainable manifest for youtube_transfer_queue.
• Repeat-safe — use_archive records downloaded IDs (per mode) and skips them on later runs; YouTube mix/radio URLs are auto-cleaned to the seed video.
• Stats-ready ledger — every completed download call appends a JSONL entry with timestamp, destinations, files, bytes, uploader, and transfer status.
• Plex playlist sync — when Plex credentials are configured, downloaded audio is added to yt-dlp Downloads by default.

MCP Tool Reference

youtube_download parameters

When Plex credentials are configured, successful downloads that produced audio files search Plex for each downloaded track, create the target playlist if needed, and add missing tracks while skipping entries already present. The default playlist is yt-dlp Downloads; set YTDLP_PLEX_PLAYLIST or pass plex_playlist to override it. Plex errors are reported as plex_playlist_error and do not make the completed download fail. JSON responses include a plex_playlist summary with matched, added, already_present, and missing counts.

Canonical metadata matching through MusicBrainz/AcoustID is documented in docs/musicbrainz-acoustid.md. youtube_download automatically runs high-confidence MusicBrainz retagging for downloaded audio when YTDLP_ACOUSTID_CLIENT_KEY is configured; youtube_identify remains available for previewing or repairing existing library files, with manual tag writes enabled by write_tags=true.

youtube_download JSON response

With response_format=json, the call returns a single object describing the batch:

Each items[].status is one of:

• ok — succeeded with files.
• partial — an error occurred but some files were still produced.
• failed — errored with no files.
• skipped — nothing new (already in the archive).

Optional keys are attached only when the relevant stage ran:

• metadata_retag — MusicBrainz/AcoustID auto-retag summary (attempted, matched, written, skipped, errors, or an error string); present when YTDLP_ACOUSTID_CLIENT_KEY is configured.
• plex_playlist — Plex playlist summary (playlist, matched, added, already_present, missing); plex_playlist_error is set instead if the Plex update failed (a Plex failure does not fail the download).
• history_error — set when the download succeeded but the JSONL ledger append failed.

youtube_probe takes urls and response_format.

youtube_plex_playlist

Build or preview Plex audio playlists from successful ytdl-rmcp download history.

Actions:

Candidates are history-derived and audio-only. Failed or retained-staging transfers are intentionally excluded.

apply can return plexamp_url, plex_web_url, and playback_link_status. plexamp_url is a best-effort generated listen.plex.tv playback link, not an official Plexamp API guarantee. The regular Plex playlist API calls use the official Plex Media Server API.

youtube_transfer_queue

List and drain server-created transfer failure manifests.

Actions:

The queue never accepts arbitrary filesystem paths. Retry uses the original target paths recorded at failure time and re-checks local target policy before transfer. Manifests are created by the local server when a transfer fails while the staging directory, manifest ID, file list, and original targets still match.

youtube_identify parameters

youtube_identify runs Chromaprint fpcalc, sends the fingerprint to AcoustID, and returns MusicBrainz recording candidates. When the best candidate is high-confidence, it also fetches the MusicBrainz recording/release data and includes a retag_preview showing the canonical artist, title, release, release date, release type, track number, and MusicBrainz IDs. By default it is preview-only. With write_tags=true, it writes the preview to the file with Lofty, including common title/artist/album/date/track fields plus MusicBrainz recording, release, release-group, and release-type tags. It requires YTDLP_ACOUSTID_CLIENT_KEY; set FPCALC_PATH if fpcalc is not on PATH.

youtube_search parameters

youtube_search_ui accepts the same input and returns the same search payload, plus MCP App metadata for hosts that can render the embedded UI. Hosts without app rendering still receive the normal search result text and structured data.

MCP App pattern

youtube_search_ui is the widget-backed entry point for this repo:

• The tool descriptor advertises _meta.ui.resourceUri.
• The server exposes ui://ytdl-rmcp/youtube-search.html through resources/list and resources/read.
• The resource uses text/html;profile=mcp-app and an explicit CSP metadata block.
• .mcpb / .dxt packaging installs the local server; UI resources are still advertised through the MCP tools/resources protocol and are host-rendered.

youtube_stats parameters

JSON stats include total_downloads, total_files, total_bytes, skipped_entries, by_kind, by_uploader, and recent. Bucket fields include downloads (compatibility alias for call count), calls, items, files, bytes, and human-readable size. Malformed ledger lines are skipped and counted instead of failing the whole stats call. If a download succeeds but the ledger append fails, the download response still succeeds and includes history_error in JSON output.

CLI Reference

The CLI owns process setup and diagnostics; media operations are exposed as MCP tools.

The npm launcher exposes the same binary:

npx -y @dinglebear/rytdl doctor
npx -y @dinglebear/rytdl serve

Install

Run the guided installer through npm:

npx -y @dinglebear/rytdl setup

Or install the command globally:

npm i -g @dinglebear/rytdl
ytdl-rmcp setup

The npm package downloads the matching GitHub Release binary during postinstall; the installed command is the Rust binary served through a tiny Node launcher. You can also use the one-line installer:

curl -fsSL https://raw.githubusercontent.com/dinglebear-ai/rytdl/main/scripts/install.sh | bash

Or download the binary tarball for your platform from Releases, or build it (see below). The guided setup fetches yt-dlp + ffmpeg, prompts for your audio/video target paths, detects which agent CLIs are present, and registers the server into the ones you pick.

Quickstart

After setup, prove the read-only path first:

npx -y @dinglebear/rytdl doctor

For raw MCP clients, call a read-only tool with JSON-RPC tools/call:

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "youtube_search",
    "arguments": {
      "query": "lcd soundsystem live",
      "limit": 3,
      "response_format": "json"
    }
  }
}

Then probe a known URL before downloading:

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "youtube_probe",
    "arguments": {
      "urls": "https://www.youtube.com/watch?v=dQw4w9WgXcQ"
    }
  }
}

Use youtube_download only after YTDLP_TARGET_PATH or YTDLP_VIDEO_TARGET_PATH points at a destination you are comfortable letting trusted MCP callers write into.

Client Configuration

Run without subcommands, npx -y @dinglebear/rytdl serves MCP over stdio. Register it yourself:

# Claude Code
claude mcp add -s user @dinglebear/rytdl -e YTDLP_TARGET_PATH=nashost:/media/music -e YTDLP_EXTRACTOR_ARGS=youtube:player_client=android -- npx -y @dinglebear/rytdl
# Codex
codex  mcp add --env YTDLP_TARGET_PATH=nashost:/media/music --env YTDLP_EXTRACTOR_ARGS=youtube:player_client=android @dinglebear/rytdl -- npx -y @dinglebear/rytdl
# Gemini CLI (command is positional, env last)
gemini mcp add -s user @dinglebear/rytdl npx -y @dinglebear/rytdl -e YTDLP_TARGET_PATH=nashost:/media/music -e YTDLP_EXTRACTOR_ARGS=youtube:player_client=android

If you already installed a standalone binary with npm i -g @dinglebear/rytdl, scripts/install.sh, or a release tarball, you can use that binary path in place of npx -y @dinglebear/rytdl.

For raw MCP JSON configs, include the required target path env var and the YouTube extractor override:

{
  "mcpServers": {
    "ytdl-rmcp": {
      "command": "npx",
      "args": ["-y", "@dinglebear/rytdl"],
      "env": {
        "YTDLP_TARGET_PATH": "nashost:/mnt/user/data/media/music/yt-dlp",
        "YTDLP_VIDEO_TARGET_PATH": "nashost:/mnt/user/data/media/movies/yt-dlp",
        "YTDLP_AUTO_UPDATE": "1",
        "YTDLP_MAX_AGE_DAYS": "1",
        "YTDLP_EXTRACTOR_ARGS": "youtube:player_client=android"
      }
    }
  }
}

The checked-in .mcp.json is also a complete raw MCP profile: it declares the same user_config keys used by the plugin/bundle manifests, supplies defaults for local gateway imports, and maps every setting into the server environment. For reliable YouTube search/probe behavior it defaults YTDLP_EXTRACTOR_ARGS to youtube:player_client=android; official music-video results frequently reject yt-dlp's default YouTube clients during metadata extraction.

Runtime Surfaces

Distribution Contract

• npm launcher — npx -y @dinglebear/rytdl downloads and runs the matching GitHub Release binary. Run without subcommands, it serves MCP over stdio; npx -y @dinglebear/rytdl setup runs the guided installer. Stable releases publish the package from GitHub Actions with npm provenance.
• Claude Code plugin — .claude-plugin/plugin.json prompts for config via userConfig; .mcp.json launches npx -y @dinglebear/rytdl, which downloads the matching GitHub Release binary through npm.
• Gemini CLI extension — gemini-extension.json; install with gemini extensions install https://github.com/dinglebear-ai/rytdl. MCP clients should prefer the npm launcher command, npx -y @dinglebear/rytdl.
• Container image — ghcr.io/dinglebear-ai/rytdl:main is published on every push to main, or build locally with docker build -t ytdl-rmcp:local .. It includes ffmpeg, fpcalc, openssh-client, rclone, and rsync. See docs/container.md for MCP and mounted-library examples.
• MCP bundle (.mcpb / .dxt) — mcpb/manifest.json defines a binary-type bundle for one-click install in MCPB-capable desktop hosts. Every main release publishes ytdl-rmcp.mcpb plus a legacy ytdl-rmcp.dxt alias; both contain the same linux + windows binaries. The bundle defaults optional config values to empty strings so Claude Desktop can install it before you fill in the destination settings. Configure at least the target path in the extension settings before downloading. Build one locally from prebuilt binaries with scripts/build-mcpb.sh (needs Node for the @anthropic-ai/mcpb CLI).

Configuration

Only one value is required for downloads: YTDLP_TARGET_PATH. YTDLP_VIDEO_TARGET_PATH is required only when video files should land somewhere different from audio. youtube_search, youtube_probe, and youtube_identify do not need a destination. The install manifests default YTDLP_EXTRACTOR_ARGS to youtube:player_client=android because that client survives common YouTube metadata blocks that make official music-video search results look unavailable.

Target path forms:

• /path/to/library — local directory. Copied by the binary's own filesystem copy (no external tool); a destination nested inside the staging source is rejected. Requires YTDLP_ALLOW_LOCAL_TARGETS=true because local paths let MCP callers choose write locations.
• host:/path/to/library — SSH target. Uses the existing rsync/scp flow and honors YTDLP_SSH_OPTS.
• remote:path/to/library — rclone target with a relative remote path. Uses rclone copy; rclone must be configured on PATH for the server process.
• rclone:remote:/path/to/library — explicit rclone target for an absolute remote path; without the rclone: prefix, host:/path remains SSH by design.

Maintainers: this table is maintained by hand. scripts/check-packaging.sh cross-checks the machine-readable config surfaces (the Claude plugin, .mcp.json user_config, gemini-extension.json, and mcpb/manifest.json) and verifies this README mentions every mapped env var name. It does not compare README descriptions or defaults. When you add, rename, or remove a YTDLP_*, FFMPEG_*, FPCALC_PATH, or YTDLP_LOG env var, update this table manually.

Authentication

This server speaks stdio only — it opens no network listener, binds no port, and has no HTTP auth layer. Access is whatever the local MCP client grants to the process it launches. External credentials live in operator-controlled environment variables, local config, SSH agents, rclone config, or Plex/AcoustID variables.

MCP callers never provide Plex tokens, SSH keys, rclone credentials, hash pins, or downloader binary paths in tool arguments. Tool arguments carry media URLs, search/probe inputs, output mode choices, and destination selectors that are validated against the process configuration.

Safety And Trust Model

Bootstrap trust model

By default, first run resolves tools in this order: explicit env override, PATH, cache, then HTTPS download from the upstream release source. Set YTDLP_SHA256 and/or FFMPEG_SHA256 to require an exact executable digest after resolution or download. These pins verify bytes on disk, but they do not fetch upstream signatures or automatically discover trusted digests; operators who need a fully pinned supply chain should provide known-good binaries through YTDLP_PATH / FFMPEG_PATH plus matching SHA-256 pins, or disable yt-dlp auto-update.

For stricter bootstrap control, combine YTDLP_PATH / FFMPEG_PATH with matching YTDLP_SHA256 / FFMPEG_SHA256 pins. Hash pins verify the resolved executable bytes; they are not upstream signature verification.

Security posture

This server is designed to run with trusted callers and operator-supplied config — it is not a hardened multi-tenant boundary.

• Tool-call URLs reach yt-dlp. Whatever urls an MCP caller passes are handed to yt-dlp, a powerful extraction tool. Only point callers at it that you trust. Tool-call URLs are validated as http/https before they reach yt-dlp, and every positional is passed after a -- end-of-options separator so a --prefixed value can't be parsed as a flag; the trust assumption above still holds regardless.
• SSH is key-only and non-interactive. SSH targets force BatchMode=yes and StrictHostKeyChecking=accept-new, so a TTY-less server fails fast instead of prompting; there is no password auth. Auth comes from your SSH key/agent and any options you add via YTDLP_SSH_OPTS.
• Target paths are validated. Local and SSH paths must be absolute and may not contain traversal or control characters. SSH remotes also reject option-like and whitespace/control-bearing values before anything reaches ssh/rsync/scp; remote paths are single-quote-escaped for the remote shell. Rclone targets must be remote:path or explicit rclone:remote:/path and reject control characters.

Requirements

• rsync is recommended for SSH transfers and falls back to scp when absent. Local transfers need no external tool — the binary copies the tree itself.
• ssh plus passwordless key-based auth when using host:/path targets.
• rclone plus a configured rclone remote when using remote:path or rclone:remote:/path targets.
• yt-dlp and ffmpeg are fetched automatically (override with YTDLP_PATH / FFMPEG_PATH, or just have them on PATH).
• youtube_identify additionally needs fpcalc; the container image includes it via libchromaprint-tools.

Development

Build and test from source:

cargo build --release                                          # Linux/macOS
cargo test && cargo clippy --all-targets -- -D warnings        # checks

# Cross-compile to Windows from Linux (needs nasm + the LLVM toolchain):
sudo apt-get install -y nasm llvm clang lld
cargo install cargo-xwin
cargo xwin build --release --target x86_64-pc-windows-msvc

On devhost/local shells, ~/.local/bin/cargo is a wrapper that can break cargo xwin; use the real rustup cargo for local Windows rehearsals:

~/.cargo/bin/cargo xwin build --release --target x86_64-pc-windows-msvc

CI (.github/workflows/) runs fmt + clippy + tests and a Windows cross-build on every push/PR, and publishes both binaries to a GitHub Release on v* tags.

This single-crate workspace uses Rust edition 2024 with MSRV 1.97.1.

Verification

Use these checks before releasing README, package, or runtime changes:

python3 /home/jmagar/workspace/soma/scripts/check-readme-guide.py README.md packages/ytdl-rmcp/README.md
npm --prefix packages/ytdl-rmcp run check
cargo fmt --check
cargo check
cargo test
scripts/check-packaging.sh
git diff --check

For a live MCP smoke, register with a temporary config and call youtube_search or youtube_probe before running youtube_download.

Architecture

Bare invocation serves MCP over stdio; setup runs the installer. A youtube_download call:

1. Resolves yt-dlp + ffmpeg (env override → PATH → cache → download) and verifies SHA-256 pins when configured.
2. Cleans mix/radio URLs, then runs yt-dlp per mode into a staging tree (staging/audio, staging/video) with metadata/thumbnail/archive flags, source metadata sidecars, and the Artist/Title [id] output template.
3. (optional) When YTDLP_ACOUSTID_CLIENT_KEY is set, fingerprints the downloaded audio and writes high-confidence MusicBrainz/AcoustID tags in-place — before transfer, so the target copy carries the canonical tags.
4. Transfers each kind's subtree to its target: local in-process copy, SSH rsync/scp, or rclone copy.
5. (optional) When Plex credentials are configured and the transfer succeeded, adds the downloaded audio tracks to the target Plex playlist.
6. Appends the completed call to the JSONL download ledger.
7. Returns a markdown or JSON summary listing files, sizes, and the actual destination(s).

See CLAUDE.md for architecture, conventions, and gotchas.

Deployment

• Local clients should use npx -y @dinglebear/rytdl or the installed rytdl binary over stdio.
• Shared runtime jobs can use the container image when ffmpeg, fpcalc, SSH, rclone, and rsync should be present without host-level installs.
• Host deployments should set destination, Plex, AcoustID, SSH, and rclone credentials in process env or host config before exposing the MCP server to callers.
• Keep YTDLP_ALLOW_LOCAL_TARGETS=0 unless trusted callers are allowed to write to local filesystem paths selected through tool arguments.

Troubleshooting

• YouTube metadata looks unavailable: set YTDLP_EXTRACTOR_ARGS=youtube:player_client=android or use the checked-in client/plugin configs, which already default it.
• Downloads succeed but transfers fail: inspect staging_kept_at or the youtube_transfer_queue list, then retry after fixing SSH, rclone, or local target permissions.
• Plex playlist updates are missing: confirm YTDLP_PLEX_URL, YTDLP_PLEX_TOKEN, and that Plex can see the transferred files.
• Tool bootstrap is too loose for your environment: pin YTDLP_PATH and FFMPEG_PATH with matching SHA-256 variables, or disable yt-dlp auto-update.

Related Servers

• soma - RMCP runtime for provider-backed MCP servers.
• unifi-rmcp - UniFi controller REST API bridge.
• tailscale-rmcp - Tailscale API bridge for devices, users, and tailnet operations.
• unraid - Unraid GraphQL bridge for NAS and server management.
• apprise-rmcp - Apprise notification fan-out bridge for many delivery backends.
• gotify-rmcp - Gotify push notification bridge for sends, messages, apps, and clients.
• arcane-rmcp - Arcane Docker management bridge for containers and related resources.
• yarr - Media-stack bridge for Sonarr, Radarr, Prowlarr, Plex, and related services.
• synapse-rmcp - Local Synapse workflow server for scout and flux actions.
• cortex - Syslog and homelab log aggregation MCP server.
• axon - RAG, crawl, scrape, extract, and semantic search project.
• labby - Homelab control plane and MCP gateway project.
• lumen - Local semantic code search MCP server.

Documentation

This README is the curated operator entry point. The generated or machine-readable source of truth lives in the package, plugin, and bundle manifests that scripts/check-packaging.sh cross-checks.

License

Original Dinglebear-authored portions of this project are licensed under AGPL-3.0-only. Separate commercial licensing is available for organizations that need terms outside the AGPL. Third-party material remains under its original license. See LICENSING.md.

Rust MCP naming pattern

This repo follows the Rust MCP server naming convention:

• Repo: dinglebear-ai/rytdl
• Cargo crate: ytdl-rmcp
• CLI alias: rytdl
• npm package: @dinglebear/rytdl