Yarr MCP

yarr

Self-hosted media fleet operations across Sonarr, Radarr, Plex, and related apps over MCP and CLI.

If you run Claude Code, Codex, or Gemini CLI against a self-hosted media stack, yarr gives an agent one consistent way to query and control all of it instead of eleven different ad hoc integrations. It is an upstream-client MCP server: it does not replace those applications or mirror every REST endpoint as a web UI. Its job is to provide one consistent tool surface for agents and one equivalent CLI surface for operators.

Not for: a general-purpose REST gateway to arbitrary services, or a scheduler/automation engine in its own right. yarr only talks to the service kinds it knows about, and only does what you or your agent ask it to do.

Contents

• Naming
• Capabilities And Boundaries
• Install
• Unraid Plugin
• 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

This repository is published at github.com/dinglebear-ai/yarr.

The Rust package and installed binary are both yarr. The npm launcher package is yarr-mcp because the shorter yarr name is occupied on npm; installing the launcher still gives you a yarr command. The MCP registry name is ai.dinglebear/yarr, and Docker images use ghcr.io/dinglebear-ai/yarr. Production Compose deployments select that image by immutable manifest digest.

Plugin naming is intentionally split:

• yarr is the full MCP server plugin. It launches the repository-coupled @dinglebear/yarr npm package over stdio and includes every per-service fallback skill; it does not commit a platform-specific binary.
• sonarr, radarr, prowlarr, overseerr, sabnzbd, qbittorrent, plex, jellyfin, tautulli, tracearr, and bazarr are skills-only plugins with no MCP server and no bundled binary.

Capabilities And Boundaries

yarr wraps a configured media automation fleet through one action-dispatched service layer. The same implementation backs MCP and CLI calls, so behavior does not drift between "agent used the tool" and "operator ran the command."

Primary capabilities:

• Fleet status checks across the configured services.
• Credentialed upstream API passthrough for known service kinds.
• Table-driven OpenAPI operation metadata for Sonarr, Radarr, Prowlarr, Overseerr, Jellyfin, and Plex. The executor preserves the declared parameter, request-media, and successful-response transport contract; unsupported rows are excluded and listed in the generated capability matrix.
• Curated commands for SABnzbd, qBittorrent, Tautulli, Bazarr, and Tracearr, whose upstreams do not ship usable machine-readable specs.
• Code Mode over MCP for multi-step media automation scripts.
• Snippet storage and execution for repeatable Code Mode workflows.
• Skills-only direct-HTTP plugin fallbacks for each individual service.

Boundaries:

• yarr does not store or schedule media jobs on its own.
• It does not expose a local REST action API or embedded web UI.
• It does not accept arbitrary unknown service kinds.
• MCP callers never provide credentials, tokens, keys, or secrets as action arguments. Credentials come from environment variables, config files, or strict per-service plugin config JSON.

Install

The native release installer is the safest default because it verifies the release archive and SHA-256 before installing yarr into ~/.local/bin:

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

The npm launcher is a coupled distribution, not an alias for whatever version happens to be tagged latest. Verify the exact release version exists on npm before using it:

YARR_VERSION=2.1.0
npm view "@dinglebear/yarr@${YARR_VERSION}" version
npx -y "@dinglebear/yarr@${YARR_VERSION}" mcp
# Or, after the same availability check:
npm i -g "@dinglebear/yarr@${YARR_VERSION}"

Never use an unpinned npx @dinglebear/yarr or @latest in an MCP manifest: npm may still point at an older launcher after a partial release. At the time of this documentation refresh, GitHub release v2.1.0 is public but @dinglebear/[email protected] is not yet available on npm; recovery is tracked in issue #80. Use the native installer, a verified release archive, a source build, or the independent Unraid package until the exact npm version resolves.

Unraid Plugin

The coordinated Unraid distribution lives under unraid-plugin/. It combines the classic .plg installer and privileged service lifecycle, an external NestJS GraphQL extension, and Vue settings/dashboard custom elements.

Install the plugin URL from Unraid's Plugins > Install Plugin page:

https://raw.githubusercontent.com/dinglebear-ai/yarr/main/unraid-plugin/yarr.plg

Fresh installs bind Yarr to loopback. LAN or custom-address binding is rejected until authentication is configured; Tailscale Serve is the supported tailnet-only option. Service credentials stay in server-side boot configuration and are never returned to the browser. See the Unraid operator and release guide for persistence, discovery, updates, rollback, uninstall retention, troubleshooting, and release gates.

Quickstart

The first-screen 30-second path is:

export YARR_SERVICES=sonarr
export YARR_SONARR_URL=http://127.0.0.1:8989
export YARR_SONARR_API_KEY=...

yarr sonarr status
yarr mcp

Then point an MCP client at the stdio command:

{
  "mcpServers": {
    "yarr": {
      "command": "yarr",
      "args": ["mcp"],
      "env": {
        "YARR_SERVICES": "sonarr",
        "YARR_SONARR_URL": "http://127.0.0.1:8989",
        "YARR_SONARR_API_KEY": "${YARR_SONARR_API_KEY}"
      }
    }
  }
}

For Claude Code plugin installs, use the marketplace commands inside a Claude Code chat session, not in a shell:

/plugin marketplace add dinglebear-ai/yarr
/plugin install yarr@yarr

The full plugin starts the exact launcher pinned in its manifest. Confirm that version exists before installing or debugging the plugin:

npm view @dinglebear/[email protected] version

Until issue #80 is resolved, the full plugin cannot start from npm. The service-specific skills-only plugins do not depend on that launcher and remain available for direct upstream workflows.

Install one skills-only plugin when you want direct service scripts without the MCP server:

/plugin install sonarr@yarr
/plugin install plex@yarr

Client Configuration

stdio

stdio is the preferred local MCP transport. It starts yarr mcp on demand and does not require binding a local HTTP port.

{
  "mcpServers": {
    "yarr": {
      "command": "yarr",
      "args": ["mcp"],
      "env": {
        "RUST_LOG": "info,yarr=debug"
      }
    }
  }
}

Streamable HTTP

Run a persistent server when several clients or machines should share one MCP endpoint:

YARR_MCP_TOKEN=change-me yarr serve

{
  "mcpServers": {
    "yarr": {
      "url": "http://127.0.0.1:40070/mcp",
      "headers": {
        "Authorization": "Bearer ${YARR_MCP_TOKEN}"
      }
    }
  }
}

HTTP MCP initialization smoke call:

curl --fail http://127.0.0.1:40070/mcp \
  -H "Authorization: Bearer $YARR_MCP_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-11-25","capabilities":{},"clientInfo":{"name":"curl-smoke","version":"1"}}}'

Use an MCP client such as mcporter for session-aware tools/list and Code Mode calls; a raw one-shot tools/call is not a complete MCP session.

Runtime Surfaces

Set YARR_MCP_TOOL_MODE=flat to advertise one action-dispatched MCP tool per configured service instead of the single Code Mode tool. That mode is useful behind a gateway such as Labby that already provides its own discovery and sandbox layer. codemode is the default and the right choice for standalone MCP clients.

MCP Tool Reference

By default, MCP exposes one tool named yarr.

Inside Code Mode, scripts can use:

• Per-service callables such as sonarr.get_series(), radarr.post_movie({ body }), prowlarr.get_indexer(), and plex.get_sessions().
• Curated commands such as qbittorrent.download_queue() and tautulli.stats_activity().
• Raw passthrough helpers at api.<service>.get/post/put/delete(path, body).
• callTool(action, params) for the underlying action-dispatch escape hatch.
• codemode.search(query) and codemode.describe(path) for discovery.
• codemode.run(name, input), codemode.snippets(), and writeArtifact(...) for reusable scripts and artifacts.

Generated callables come from metadata tables. They preserve method/path, required path arguments, known query names, JSON-body presence, and a bounded single-file multipart escape hatch. They do not enforce every required query, header/cookie parameter, style/explode rule, form schema, media type, or response schema. See docs/API.md before relying on a generated operation for a non-JSON or serialization-sensitive endpoint.

Example:

async () => {
  const queue = await radarr.get_queue();
  await radarr.post_command({
    body: { name: "MoviesSearch", movieIds: [456] }
  });
  return { queued: queue.records?.length };
}

Generic Actions

These actions work for every configured service kind:

Code Mode Actions

There is no confirm argument. CLI destructive commands dispatch immediately. MCP direct and nested Code Mode destructive calls require elicitation and fail closed if the peer cannot elicit or approval is not granted.

CLI Reference

The CLI is service-grouped:

yarr help
yarr radarr status
yarr sonarr get --path /api/v3/system/status
yarr radarr post --path /api/v3/command --body '{"name":"RefreshMovie"}'
yarr sonarr put --path /api/v3/series/editor --body '{"seriesIds":[1],"qualityProfileId":4}'
yarr radarr delete --path /api/v3/movie/12

# Generated operations for spec-backed services.
yarr sonarr op get_series
yarr radarr op post_command --args '{"body":{"name":"MoviesSearch","movieIds":[456]}}'

# Curated commands for doc-only services.
yarr qbittorrent queue
yarr tautulli activity

# Code Mode and snippets.
yarr codemode --code 'async () => sonarr.get_system_status()'
yarr snippet list

There is no --service flag. Infra commands such as help, codemode, and snippet are service-less.

Configuration

Copy .env.example or use config.example.toml as a starting point. Common environment variables:

YARR_MCP_HOST=127.0.0.1
YARR_MCP_PORT=40070
YARR_MCP_TOKEN=change-me

YARR_SERVICES=sonarr,radarr,prowlarr,tautulli,overseerr,bazarr,tracearr,sabnzbd,qbittorrent,plex,jellyfin
YARR_SONARR_URL=http://sonarr:8989
YARR_SONARR_API_KEY=...
YARR_RADARR_URL=http://radarr:7878
YARR_RADARR_API_KEY=...
YARR_QBITTORRENT_URL=http://qbittorrent:8080
YARR_QBITTORRENT_USERNAME=...
YARR_QBITTORRENT_PASSWORD=...
YARR_PLEX_URL=http://plex:32400
YARR_PLEX_TOKEN=...

Supported service kinds are sonarr, radarr, prowlarr, tautulli, overseerr, bazarr, tracearr, sabnzbd, qbittorrent, plex, and jellyfin.

*_API_KEY covers most Arr-style services. qBittorrent uses username/password login. Plex and Jellyfin token headers are handled separately.

YARR_MCP_TOOL_MODE=codemode is the default. Use YARR_MCP_TOOL_MODE=flat only when a gateway should see separate per-service tools.

Authentication

YARR_MCP_TOKEN authenticates /mcp on any HTTP bind, including loopback. It receives the scopes in YARR_MCP_STATIC_TOKEN_SCOPES, which defaults to read-only yarr:read. A read-only bearer deployment must use YARR_MCP_TOOL_MODE=flat; using the default Code Mode surface requires an explicit yarr:write scope. On 127.0.0.1 or localhost with no explicit bearer/OAuth configuration, auth is bypassed for local development.

Auth states:

Unauthenticated health endpoints are /health, /ready, /status, and /metrics. /status redacts secrets; /metrics exposes HTTP and bounded domain metrics and should be network-restricted when needed.

Safety And Trust Model

Secrets stay server-side. MCP clients provide action parameters, never upstream tokens. Query-string secrets such as apikey=, token=, and X-Plex-Token are rejected by path validation.

help is public at the action layer, but mounted HTTP transports still require bearer or OAuth transport auth. service_status requires yarr:read. Credentialed passthrough, generated operations, curated write operations, and Code Mode require yarr:write; write satisfies read.

Generated DELETE operations, api_delete, download_remove, stats_delete_image_cache, and trace_terminate_stream are destructive. CLI commands dispatch them immediately. MCP callers get an interactive elicitation prompt at the actual dispatch point, including inside Code Mode, with no call argument or nested callTool path that can skip it.

Responses are capped by the shared token-limit layer before they are returned to MCP clients.

Architecture

MCP shim (src/mcp/tools.rs)   CLI shim (src/cli.rs)
  JSON args -> dispatch          argv -> dispatch
        \                          /
         execute_service_action (src/actions/dispatch.rs)
             shared validation + curated-command dispatch
                          |
         YarrService (src/app.rs + src/app/*.rs)
           validation, service lookup, response shaping
                          |
         YarrClient (src/yarr.rs)
           network calls and auth headers

src/mcp.rs and src/cli.rs are thin facades re-exporting their own submodules. execute_service_action makes CLI-to-MCP parity structural rather than something kept in sync by hand.

The service layer owns:

• supported service catalog and config lookup
• safe path validation
• credential redaction
• qBittorrent login flow
• response normalization

Distribution Contract

The source of truth for release identity is the version shared by Cargo.toml, Cargo.lock, xtask/Cargo.toml, .release-please-manifest.json, packages/yarr-mcp/package.json, and server.json.

Generated code and docs must be regenerated from the committed source inputs, not patched by hand:

• Generated OpenAPI operations live under src/openapi/generated/ and come from vendored specs in specs/.
• Curated actions live in the handwritten action registries and docs.
• Plugin manifests stay versionless; marketplaces derive plugin version from the git commit SHA.
• The npm package version and the GitHub Release tag must match.
• server.json must name the exact yarr-mcp npm version and stdio launch contract under the ai.dinglebear/yarr registry identity.
• The Docker image path is ghcr.io/dinglebear-ai/yarr; production deployment uses a promoted immutable @sha256: digest.

Development

cargo run -- help
cargo fmt --check
cargo test
cargo clippy -- -D warnings
cargo build --release

When changing generated operations:

cargo xtask gen-openapi
cargo xtask tool-docs
cargo test --test parity

When changing plugin packaging:

cargo test --test plugin_contract
cargo test --test template_invariants

Verification

Use the README guide checker before landing documentation changes:

python3 /home/jmagar/workspace/soma/scripts/check-readme-guide.py README.md

Use the package and Rust checks for distribution-sensitive work:

npm --prefix packages/yarr-mcp run check
cargo fmt --check
cargo check
cargo test
git diff --check

For live install verification, validate the native installer and only exercise the npm path after the exact coupled version resolves:

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

YARR_VERSION=2.1.0
npm view "@dinglebear/yarr@${YARR_VERSION}" version
npx -y "@dinglebear/yarr@${YARR_VERSION}" mcp

Documentation changes also require the repository-local link and anchor guard:

python3 scripts/check-doc-links.py

Deployment

Run as a persistent HTTP MCP server:

YARR_MCP_HOST=0.0.0.0 \
YARR_MCP_PORT=40070 \
YARR_MCP_TOKEN=change-me \
YARR_MCP_STATIC_TOKEN_SCOPES=yarr:read \
YARR_MCP_TOOL_MODE=flat \
yarr serve

Docker Compose deployments are covered by docker-compose.yml, docker-compose.prod.yml, docs/DOCKER.md, and docs/DEPLOYMENT.md. Production deployments should put yarr behind a trusted reverse proxy or MCP gateway when exposed outside loopback.

Troubleshooting

• 401 or 403 from HTTP MCP: confirm YARR_MCP_TOKEN, OAuth mode, and gateway headers.

• unknown service: confirm the service is listed in YARR_SERVICES and has a matching YARR_<SERVICE>_URL.

• upstream 401: confirm the service-specific API key or token in the server environment, not in tool arguments.

• query-string secret rejected: remove tokens from --path and put them in config.

• plugin skill cannot reach a service: its per-service config JSON under ~/.config/lab-<service>/config.json is missing or stale. It is written by the plugin's SessionStart / ConfigChange hook, so start a new session (or change a plugin setting) to refresh it. To rebuild it by hand, export the settings yourself — the scripts read CLAUDE_PLUGIN_OPTION_*, which only the hook sets:

  CLAUDE_PLUGIN_OPTION_SONARR_URL=https://sonarr.example.com \
  CLAUDE_PLUGIN_OPTION_SONARR_API_KEY=… \
    bash plugins/sonarr/scripts/setup.sh
  

  Skill scripts run through the Bash tool, which receives no CLAUDE_PLUGIN_OPTION_* variables — only hook processes do. That is why these plugins need the hook: it is the one channel that can carry a sensitive: true value to a skill, since sensitive values are never substituted into skill prose and these plugins have no MCP server to receive an ${user_config.*} env block.

• Code Mode cannot find a callable: use codemode.search(...) and codemode.describe(...); generated names follow upstream OpenAPI operation IDs after normalization.

• npm or the full plugin reports E404/ETARGET: check the exact pinned launcher with npm view @dinglebear/[email protected] version. Do not fall back to unpinned latest; use the native binary while issue #80 is open.

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-rmcp - 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.
• ytdl-rmcp - Media download and metadata workflow server.
• 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

The source of truth docs split is:

• docs/README.md is the audience-oriented documentation index.
• docs/DOCS.md defines authority, generated-doc rules, freshness checks, and where new documentation belongs.
• docs/API.md covers action contracts and Code Mode call shape.
• docs/CONFIG.md, docs/ENV.md, and docs/AUTH.md cover configuration, environment variables, auth states, and tool modes.
• docs/QUICKSTART.md covers native, source, exact-version npm, and Unraid first-run paths.
• docs/MCP_SCHEMA.md and docs/TOOLS_ACTIONS_ENDPOINTS.md are checked schema/action references.
• docs/DEPLOYMENT.md, docs/DOCKER.md, and docs/SYSTEMD.md cover production runtime choices.
• docs/CI.md, docs/SCRIPTS.md, and docs/runbooks/ cover validation, releases, incidents, and rollback.
• docs/PATTERNS.md holds conventions shared across the RMCP server family.
• docs/PLUGINS.md, plugins/README.md, and plugins/yarr/README.md cover marketplace packaging and launcher availability.
• unraid-plugin/README.md is the operator, API, recovery, verification, and release guide for the classic Unraid distribution.
• CLAUDE.md holds repo-local agent memory and the "How to add an action" checklist.

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.