A workflow MCP server over a self-hosted Jellyseerr / Overseerr instance — the request layer that fronts Jellyfin/Plex/Emby plus Radarr and Sonarr. The unit of work is not "download a movie"; it is search → resolve the exact TMDB-backed title → check availability and request state → create a guarded request that Radarr/Sonarr act on. Jellyseerr owns permissions, quotas, routing, and status; this server never touches Radarr/Sonarr directly.

Two properties make it safe to hand an agent:

• Guarded writes. The one mutating tool (seerr_request_media) defaults to mode: preview — it resolves the title and returns the exact payload that would be submitted without writing anything. The real request fires only on mode: request, which asks for an explicit confirmation and submits only once that confirmation comes back accepted.
• Operator-safe output. Every payload passes through one normalization layer (src/services/seerr/normalizers.ts) that allow-lists the fields it emits, so operator email, Plex/Jellyfin tokens and media-server IDs, and internal serviceUrl hosts never reach the model — requesters are projected to { id, displayName }. Redaction is always on; filesystem paths are the single opt-in, behind includePaths.

Tools

Six tools covering the request workflow — discover (search) → confirm (get) → understand routing (service_options) → request (request_media) → track (request_status / list_requests):

Every status field is decoded to { raw, label } — both the numeric code Jellyseerr returns and a human label — so an agent never has to hardcode the enum mapping.

seerr_search_media

Title disambiguation entry point. Wraps GET /search, filters to movies and TV (people are always excluded), and decodes availability when the title is tracked.

• Free-text title queries matched against TMDB; movie / tv / all media-type filter
• Decoded availability (status, plus status4k when 4K is enabled) for tracked titles only
• Pagination by page, with a per-call result limit to cap output size
• Optional ISO 639-1 language for localized titles/overviews
• Empty results are a normal success — returns [] with a guidance notice, not an error

seerr_get_media

Confirm the exact title before a write. Wraps GET /movie/{id} or GET /tv/{id}, optionally a season's episodes.

• Availability plus any existing open request for the title (avoids duplicate requests)
• TV: omit seasonNumber for a per-season summary, or pass one to fetch that season's episode list (season 0 is Specials)
• A TMDB ID that doesn't resolve surfaces as a clean media_not_found with a search-recovery hint (Jellyseerr's raw HTTP 500 is classified in the service layer)

seerr_list_requests

Review recent requests and their lifecycle. Wraps GET /request.

• Lifecycle filter (pending, processing, available, failed, …), mediaType, and requestedById filters
• Sort by created (added) or last-changed (modified), ascending or descending
• take / skip pagination; the enrichment trailer echoes the filter set the server applied
• Requester is PII-redacted to { id, displayName }
• Titles aren't on request objects. includeTitles: true joins them from the media records — one lookup per distinct title on the page, so a 4K and a non-4K request for one film cost one call. Default is off, keeping the call to a single upstream read. Rows that can't be resolved keep every other field and are disclosed in the notice

seerr_request_media

The only mutation in the surface, and it is triple-guarded:

1. mode: preview (default) resolves the title and returns the exact POST /request payload that would be submitted — no write. A sloppy call shows the payload and changes nothing.
2. mode: request asks for an explicit confirmation and submits only once it comes back accepted. Declining, cancelling, or answering with anything the confirmation schema rejects cancels before submission (request_cancelled) — there is no branch that proceeds without an acceptance, so a client that never answers leaves the write un-run.
3. destructiveHint: true surfaces the risk in client-side approval flows.

Over HTTP this needs MCP_SESSION_MODE to resolve to stateful (the default auto does): 2025-era clients answer the confirmation over a live session, which stateless has no way to hold open.

• Capability validation (4K enabled? seasons valid? partial requests allowed?) runs locally against cached instance settings before any POST, so a bad request fails with an actionable typed error instead of a failed write
• TV requests take seasons: "all" or an explicit list (e.g. [1, 2]); Specials are excluded unless the instance enables them
• Optional routing overrides (serverId, profileId, rootFolder, languageProfileId) — omit to use Jellyseerr's defaults (recommended)
• An existing request for the title is surfaced in the output; a duplicate rejection from Jellyseerr maps to a typed duplicate_request pointing back at it

seerr_service_options

Lets an agent reason about request capability and routing without a separate status tool. Fans out service + settings + version reads with Promise.allSettled, so one failed leg degrades to a disclosed notice rather than failing the call.

• Instance capability summary: Jellyseerr version, media server, and the movie4kEnabled / series4kEnabled / partialRequestsEnabled / specialEpisodesEnabled flags
• Per-service routing: server ID, default-server flag, 4K capability, and the active + available quality profiles (IDs and names, safe to surface)
• Filesystem root-folder paths and free space are operator-private — omitted unless includePaths: true

Resource and prompt

All request data is also reachable via tools — request enumeration is the job of seerr_list_requests (filterable, the tool-only access path), so the collection is intentionally not exposed as a resource. There are no prompts; the guarded-write workflow lives in the tool, not a prompt template.

Features

Built on @cyanheads/mcp-ts-core: