arr-py-client 0.9.2

Typed Python client + MCP server + declarative config sync for Radarr, Sonarr, and Prowlarr

arr-py-client

Typed Python client + MCP server + declarative config sync for Radarr, Sonarr, and Prowlarr. One SDK for scripts, one server for LLM agents, one YAML pipeline for "make my stack match this file."

• Highlights
• Install
• Quickstart
  • SDK
  • MCP server (stdio)
  • MCP server (Docker)
• Features in depth
• How it compares
• Documentation
• Development
• License

Highlights

Fully typed, fully async-ready SDK. Pydantic v2 models, httpx transport, complete Radarr v3 + Sonarr v3 + Prowlarr v1 endpoint coverage, sync and async mirrors of every operation. Ships py.typed so mypy / basedpyright / Pylance pick it up immediately.

Drop-in MCP server for Claude, Cursor, and custom agents. arr-py-mcp exposes 60+ tools over stdio or streamable-HTTP — the LLM can list movies, explain a release grade, run a queue janitor, or sync declarative config without any bespoke tool plumbing on your side. Every response carries _meta.action_hints so chains self-suggest their next call.

Multi-tenant out of the box. One environment-var provider for solo use; a ClientProvider + OAuth 2.1 path for SaaS deployments where every user has their own stack. Per-tool scopes, RFC 6750 WWW-Authenticate headers, and an audit= hook are already wired.

Declarative config sync (Recyclarr-style, in Python). Describe tags, custom formats, and quality profiles in YAML (or JSON / TOML). Call plan() to preview, apply() to converge — same desired-state model for Radarr and Sonarr.

Workflow primitives that are boring on purpose. queue.janitor() (policy-based cleanup with named bundles), library.backfill() (rate-limited missing-content search with .estimate()), releases.explain() (human-readable grading + .advice()), arr_health() (cross-brand rollup). No LLM required — they're ordinary Python you can schedule from cron.

Webhook receivers. Typed events from Radarr / Sonarr, plumbed through FastAPI or plain WSGI with zero extra deps.

Testing utilities. make_fake_radarr() / make_fake_sonarr() for in-process fakes and an @replay(...) decorator for record-on-miss / replay-on-hit against real instances.

Small, focused CLI. arr-py handles status and basic add; workflows deliberately live in Python + MCP, not shell.

Python 3.11 – 3.14.

Install

pip install arr-py-client              # core SDK
pip install arr-py-client[mcp]         # + MCP server
pip install arr-py-client[config]      # + YAML loader for config_sync
pip install arr-py-client[webhooks]    # + FastAPI receiver helper

Works with pip, uv, pipx, and any PEP 621 installer.

Quickstart

SDK

from arr_py_client import RadarrClient

with RadarrClient(base_url="http://radarr:7878", api_key="YOUR_KEY") as client:
    for m in client.movies.list()[:5]:
        print(m.id, m.title, m.year)

Or let it read RADARR_BASE_URL / RADARR_API_KEY from the environment (or a .env file):

from arr_py_client import RadarrClient

with RadarrClient() as client:
    print(len(client.movies.list()))

Async twins ship for every client — AsyncRadarrClient, AsyncSonarrClient, AsyncProwlarrClient:

import asyncio
from arr_py_client import AsyncSonarrClient

async def main() -> None:
    async with AsyncSonarrClient() as client:
        series = await client.series.list()
        print(len(series), "shows")

asyncio.run(main())

MCP server (stdio)

Point Claude Desktop, Cursor, or any MCP client at a local process:

pip install arr-py-client[mcp]

export RADARR_BASE_URL=http://radarr:7878 RADARR_API_KEY=...
export SONARR_BASE_URL=http://sonarr:8989 SONARR_API_KEY=...
export PROWLARR_BASE_URL=http://prowlarr:9696 PROWLARR_API_KEY=...  # optional

arr-py-mcp                      # stdio — register with Claude / Cursor
arr-py-mcp --transport http     # streamable-http for remote clients

MCP server (Docker)

Prebuilt image on every release. Smoke-test in one command:

docker run --rm -p 3000:3000 \
  -e RADARR_BASE_URL -e RADARR_API_KEY \
  -e SONARR_BASE_URL -e SONARR_API_KEY \
  -e PROWLARR_BASE_URL -e PROWLARR_API_KEY \
  ghcr.io/allada-homelab/arr-py-client:latest

Or use the bundled compose file (reads .env from the repo root):

just mcp-up                     # docker compose up -d
curl -sSL -X POST http://localhost:3000/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"arr_health","arguments":{}}}'

Tags: latest, {version}, {major}.{minor}, {major}, plus sha-<short> for immutable pinning. Full walk-through: docs/guides/docker-deployment.md.

Features in depth

Typed client

Every request and response is a pydantic v2 model. Fields unknown to the client are preserved — a schema drift on the server side won't crash your code. Retries, url_base prefixes, custom httpx clients, and per-call timeouts are all first-class.

from arr_py_client import RadarrClient

with RadarrClient(base_url="http://radarr:7878", api_key="k") as client:
    movie = client.movies.get(id=42)      # pydantic model, not dict
    movie.monitored = True
    client.movies.put(id=42, body=movie)

MCP server

60+ tools, one of three transports (stdio, streamable-HTTP, SSE for legacy clients). Tools split into three groups:

• Per-brand — radarr_*, sonarr_*, prowlarr_* mirror the HTTP API surface (list / get / system status / queue / blocklist / logs).
• Composed — arr_* cross-brand workflows: arr_health, arr_sync_plan, arr_sync_apply, arr_backfill, arr_janitor_run, arr_explain_grab.
• Describe / reference — arr_describe_fields, arr_list_instances, arr_list_resources so the LLM can discover the schema at call time.

Every list / get tool returns {"data": ..., "_meta": {...}} with action_hints pointing at the next plausible tool call — LLM chains self-navigate without a giant system prompt.

Multi-tenant? Implement a ClientProvider, wire a TokenVerifier, and mount the resulting ASGI app under /mcp:

from fastapi import FastAPI
from arr_py_client.mcp import (
    build_server, CallbackTokenVerifier, InMemoryCachedProvider,
)

class MyProvider(InMemoryCachedProvider):
    async def identity(self, ctx):
        return ctx.principal.id

    async def build_client(self, ctx, brand, instance_id, identity):
        # lookup encrypted creds in your DB, return AsyncRadarrClient(...)
        ...

mcp = build_server(
    provider=MyProvider(),
    token_verifier=CallbackTokenVerifier(verify=app_auth.to_principal),
    auth=AuthSettings(issuer_url=..., resource_server_url=...),
)

app = FastAPI()
app.mount("/mcp", mcp.streamable_http_app())

Full guide: docs/guides/mcp-multi-tenant.md.

Declarative config sync

Describe the desired state once; plan tells you what would change; apply converges. Works for tags, custom formats, quality profiles.

# config.yaml
tags: [4k, anime, kids]
custom_formats:
  - name: x265
    specifications:
      - name: x265
        implementation: ReleaseTitleSpecification
        required: true
        fields: [{ name: value, value: "(h|x).?265" }]
quality_profiles:
  - name: HD-Bluray
    upgradeAllowed: true
    cutoff: 7
    formatItems:
      - { name: x265, score: -10000 }

from arr_py_client import RadarrClient
from arr_py_client.config_sync import load, plan, apply

with RadarrClient() as client:
    desired = load("config.yaml")
    plan_ = plan(client, desired)
    print(plan_.summary())
    apply(client, plan_, dry_run=False)

Full schema + more examples: docs/examples/config-sync/.

Queue janitor

Named policy bundles for common opinions; BYO policies for the rest.

from arr_py_client import RadarrClient, POLICIES

with RadarrClient() as client:
    report = client.queue.janitor(
        policies=POLICIES.default,     # .conservative | .aggressive | .ratio_preserving
        protected_trackers=("private-tracker.example",),
        dry_run=False,
    )
    print(report.total_matches)

Webhooks

Parse-only:

from arr_py_client.webhooks import parse_event, OnGrab

event = parse_event(request.json())
if isinstance(event, OnGrab):
    notify(f"Grabbed {event.movie.title if event.movie else '?'}")

FastAPI:

from fastapi import FastAPI
from arr_py_client.webhooks import fastapi_router

app = FastAPI()
app.include_router(fastapi_router(on_event), prefix="/webhooks/arr")

Zero-dep WSGI:

from wsgiref.simple_server import make_server
from arr_py_client.webhooks import wsgi_app

make_server("0.0.0.0", 9000, wsgi_app(on_event)).serve_forever()  # noqa: S104

How it compares

	arr-py-client	pyarr	Recyclarr
Pydantic v2 models	yes	no (dicts)	n/a
Async	yes	no	n/a
Radarr / Sonarr v3 coverage	yes	yes	partial (config only)
Prowlarr v1 coverage	yes	yes	yes
Lidarr / Readarr	planned	yes	yes
MCP server	yes (60+ tools)	no	no
Declarative config sync	yes (YAML/JSON/TOML)	no	yes
Queue janitor / backfill / release explain	yes	no	no
Webhook receiver helper	yes	no	no
Ships py.typed	yes	no	n/a

Documentation

• API reference: https://allada-homelab.github.io/arr-py-client/
• Architecture: docs/architecture.md
• Docker deployment: docs/guides/docker-deployment.md
• Multi-tenant MCP: docs/guides/mcp-multi-tenant.md
• TRaSH-Guides integration: docs/guides/trash-guides.md
• Connection diagnostics: docs/guides/connection-diagnostics.md
• Operator agent: docs/guides/operator-agent.md
• Roadmap: docs/roadmap.md

Development

git clone https://github.com/allada-homelab/arr-py-client
cd arr-py-client
uv sync --all-extras --all-groups
just test           # unit suite
just typecheck      # basedpyright in strict mode
just lint fmt       # ruff check + format

Integration tests (spin up real Radarr / Sonarr / Prowlarr containers via compose.integration.yml):

just int-up         # start the test containers
just test-int       # @pytest.mark.integration
just int-down       # tear down

Regenerate clients from upstream OpenAPI specs:

just gen-radarr <radarr-tag>
just gen-sonarr <sonarr-tag>
just gen-prowlarr <prowlarr-tag>

Full contributor workflow, PR conventions, and release process: CONTRIBUTING.md.

License

MIT. See LICENSE.