memby/server

Memby gateway

A small Go service that sits between the Memby Android TV client and Emby. It owns authentication, caching, search and the shaping of TV screens, so the client can stay a thin renderer.

Video never passes through here. /v1/items/{id}/playback returns a direct-play URL pointing at Emby itself; only metadata and artwork traverse the gateway.

Run it

cp .env.example .env      # from the repo root
$EDITOR .env              # set MEMBY_EMBY_URL and POSTGRES_PASSWORD
docker compose up -d --build
curl localhost:8080/readyz

Then build the TV app against it:

.\gradlew.bat assembleDebug -Pmemby.gatewayUrl=http://<host>:8080

Leaving memby.gatewayUrl blank keeps the app on its original direct-to-Emby path, so a gateway outage is one rebuild away from being routed around.

Local development

go build ./...
go test ./...
go run ./cmd/memby-server     # needs Postgres + Redis reachable

On Windows, go may need -buildvcs=false when the working tree has no usable .git.

API

All /v1 routes need Authorization: Bearer <token> from /v1/auth/login. Image URLs accept ?t=<token> instead, because the client's image loader fetches plain URLs with no headers attached.

Method	Path	Purpose
POST	/v1/auth/login	Emby credentials in, gateway token out
POST	/v1/auth/logout	Retire this device's token
GET	/v1/auth/session	Confirm a stored token is still valid
GET	/v1/home?limit=	Every launcher row in one response
GET	/v1/recommendations?refresh=1	Recommendation rows alone; refresh forces a rebuild
GET	/v1/screensaver?limit=	Backdrop pool, cached and shuffled per request
GET	/v1/search?q=&limit=	Library search
GET	/v1/items/{id}	Full metadata for one item
GET	/v1/items/{id}/playback	Resolves series → episode, returns a direct-play URL
GET	/v1/items/{id}/trailer	First local trailer, or 404
POST	/v1/items/{id}/favorite	{"value":true}
POST	/v1/items/{id}/played	{"value":true}
POST	/v1/playback/{started|progress|stopped}	Progress reporting
POST	/v1/analytics/rows	Batched row engagement from a TV
GET	/v1/images/{itemId}/{backdrop|primary|logo|thumb}	Artwork proxy
GET	/healthz, /readyz	Liveness, readiness

Server-driven rows

/v1/home returns a rows array — order, titles and kinds all decided here — plus the four fixed rows repeated flat for the client's offline cache:

{
  "rows": [
    {"id": "continue",      "title": "Continue Watching", "kind": "continue",    "items": [...]},
    {"id": "next-up",       "title": "Next Up",           "kind": "nextup",      "items": [...]},
    {"id": "favorites",     "title": "Favourites",        "kind": "favorites",   "items": [...]},
    {"id": "latest-movies", "title": "Recently Added Movies", "kind": "latest",  "items": [...]},
    {"id": "similar:sev",   "title": "Because you watched Severance", "kind": "similar", "items": [...]},
    {"id": "recommended",   "title": "Recommended from your watching history", "kind": "recommended", "items": [...]}
  ],
  "continueWatching": [...], "nextUp": [...], "favorites": [...], "latestMovies": [...],
  "partial": false
}

The TV renders whatever arrives, so a new row ships without an app release. kind picks the card shape; an unrecognised kind falls back to poster cards rather than being dropped. The user's own section toggles still hide the four fixed rows, but never rows the server invented — nobody opted out of a row that did not exist when they last opened Settings.

Recommendations

internal/recommend builds rows from viewing history. Two kinds:

• "Because you watched X" — Emby's own /Items/{id}/Similar for the most recent distinct titles, filtered down to what the user has not seen. Emby's similarity ranking is better than anything worth reimplementing here; this only removes the already-watched.
• "Recommended from your watching history" — genre and studio affinity. History is weighted by recency (0.94 per position, so the 12th item counts about half the most recent), favourites add a smaller fixed weight, and candidates are unplayed titles in the top three genres scored by affinity + a mild community-rating nudge. Titles tagged with many genres get a sqrt(n) penalty so genre-stuffing cannot buy a top slot.

Rows shorter than four items are dropped, and a user with no history gets no rows at all rather than a strip of noise.

The home screen never waits on the engine. Rows live in their own r:<userId>:rows cache key with a long TTL (2h). A cache miss serves home immediately without them and triggers a background rebuild — deduplicated per user, so four TVs waking together do the work once. Because the key sits outside the u: namespace, a favourite toggle does not throw the recommendations away; only a finished playback does, since that is the one event that genuinely changes viewing history.

The scoring is pure and unit-tested (profile_test.go), and the row assembly runs against a fake Emby (engine_test.go), so neither needs a server to verify.

Item payloads are Emby's own JSON, forwarded verbatim. That is deliberate: the Android client already models this shape, so there is no second schema to keep in sync. app/src/test/.../GatewayPayloadTest.kt and internal/api/api_test.go pin the envelope around it from both sides.

Admin interface

http://<host>:8080/admin/ — a single self-contained page for library imports, the maintenance switch and row engagement. Set MEMBY_ADMIN_TOKEN to enable it; unset, every /admin route 404s so it cannot be left exposed by accident. Paste the token into the field at the top of the page; it is kept in the browser's local storage and sent as a bearer header. Put the whole path behind your reverse proxy's own auth as well if the gateway is reachable from outside the LAN.

Method	Path	Purpose
GET	/admin/	The page
GET	/admin/api/status	Library counts, sync history, maintenance state
POST	/admin/api/sync	{"kind":"full"} or {"kind":"incremental"}
POST	/admin/api/maintenance	{"enabled":true,"message":"…"}
GET	/admin/api/analytics?days=7	Row engagement

Library import

internal/library copies Emby's catalogue into Postgres so the gateway answers from its own data instead of asking Emby per request.

• Full — pages through everything (500 items per request), then deletes any row it did not touch, which is how removals propagate. Run once to seed; re-run after reorganising the library.
• Incremental — asks Emby only for items changed since the last successful run (MinDateLastSaved, with a minute of overlap so nothing falls between runs). This is the hourly job: new episodes appear within the hour, and the weekly film drop rides along with no extra configuration.

An incremental run with no previous success upgrades itself to a full one, so a fresh deployment self-seeds on its first tick. Only one import runs at a time; the scheduler skips its tick if one is still going, and interrupted runs are marked failed at boot rather than sitting on "running" forever.

Credentials. Imports use MEMBY_SYNC_USER_ID + MEMBY_SYNC_API_KEY when set, and otherwise borrow the most recently active TV session. The fallback means a new deployment imports as soon as somebody signs in, but it stops working if that user is deleted — set a service account for anything long-lived.

What is not imported: every query runs with EnableUserData=false. Watched flags, favourites and resume positions are per-user and cannot be shared across a household, so they still come from Emby live. The imported copy powers search and the recommendation candidate pool.

Maintenance mode

Takes Memby down independently of Emby: all /v1 routes answer 503 with {"maintenance": true, "message": "…"}, and the TV shows the operator's message instead of a network error. /healthz, /readyz and /admin stay up — they are what you need while the app is deliberately off.

The switch lives in Postgres, not memory, so a restart cannot quietly bring the app back up mid-repair. Each instance caches it and re-reads every 30 seconds, so toggling it directly in the database works too.

Row analytics

The TV reports three signals per row — impression (drawn), focus (the remote landed there, with dwell), select (something was opened) — batched and uploaded every 20 seconds to POST /v1/analytics/rows. Dwell below 400 ms is dropped client-side as D-pad travel rather than attention, and the server clamps anything over 30 minutes.

Read it at /admin/, sorted by dwell. Dwell is the number worth watching: impressions only say a row was on screen, while dwell says someone stopped there. It is the fastest way to tell whether "Recommended from your watching history" is earning its slot.

Raw events are pruned after MEMBY_ANALYTICS_RETENTION (90 days) and aggregates are computed at read time, so nothing survives the prune. This is tuning telemetry, not a record of what anyone watched.

Caching

Redis holds everything user-scoped under u:<embyUserId>:*, plus session lookups under sess:<tokenHash> and recommendation rows under r:<embyUserId>:rows. Any mutation — favourite, watched, playback stopped — drops the u: keys, so the next home request re-reads Emby rather than serving a row it just contradicted. Partial home payloads are served but never cached. The r: namespace is deliberately excluded from that wipe (see Recommendations above).

Postgres holds only sessions. It is the durable half: losing Redis costs a cold cache, losing Postgres signs everyone out.

Configuration

Variable	Default	Notes
MEMBY_EMBY_URL	required	How the gateway reaches Emby
MEMBY_EMBY_PUBLIC_URL	= MEMBY_EMBY_URL	What TVs stream from
MEMBY_DATABASE_URL	required	Postgres DSN
MEMBY_REDIS_URL	redis://localhost:6379/0
MEMBY_LISTEN_ADDR	:8080
MEMBY_CLIENT_NAME	Memby	Shown in Emby's device list
MEMBY_HOME_TTL	60s	Also MEMBY_ITEM_TTL, MEMBY_SEARCH_TTL, MEMBY_SCREENSAVER_TTL
MEMBY_RECOMMEND_TTL	2h	How long computed recommendation rows stay warm
MEMBY_RECOMMEND_TIMEOUT	60s	Bounds a background rebuild
MEMBY_ADMIN_TOKEN	empty	Enables /admin. Empty = admin disabled
MEMBY_SYNC_INTERVAL	1h	Incremental import cadence; 0 disables
MEMBY_SYNC_TIMEOUT	30m	Bounds one import
MEMBY_SYNC_ON_START	false	Import at boot
MEMBY_SYNC_USER_ID / MEMBY_SYNC_API_KEY	empty	Emby service account for imports
MEMBY_ANALYTICS_RETENTION	2160h (90d)	Raw row events are pruned past this
MEMBY_SESSION_CACHE_TTL	5m	How long a token lookup stays in Redis
MEMBY_SESSION_IDLE_EXPIRY	2160h (90d)	Unused tokens are swept every 6h
MEMBY_UPSTREAM_TIMEOUT	20s

Security notes

• The sessions table stores live Emby access tokens in plaintext. Gateway tokens are stored only as SHA-256 hashes, so a database dump does not yield working gateway credentials — but it does yield working Emby ones. Treat the Postgres volume as a secret store, and encrypting emby_token at rest is the obvious next hardening step.
• Image URLs carry the gateway token in a query string, so it will appear in any access log in front of this service. That token is revocable and grants nothing outside Memby, which is why the artwork proxy exists at all.
• Nothing here terminates TLS. Put it behind your existing reverse proxy before exposing it beyond the LAN.

Not built yet

• Live change feed. Imports are polled hourly rather than driven by Emby's WebSocket, so a brand-new episode can be up to an hour late. Good enough for a household; the WebSocket would make it instant.
• Cache warming. Rows go cold after MEMBY_HOME_TTL; the first TV to ask pays for the refresh. A background refresher per active session would hide that.
• Rate limiting on /v1/auth/login.
• Per-user analytics breakdown. Events carry a user id, but the admin page only shows totals per row.