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/yarr@2.1.0 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/yarr@2.1.0 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. .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

| `s