Upnext MCP Server

upnext

One queue over every audio source — including the ones you don't control.

A library your product imports so that whatever is driving — a model, a person clicking, a script — can control audio without knowing whether the sound is coming from Spotify, a browser tab, a podcast feed, or a file on disk.

const runtime = new Runtime({ adapters: [spotify, browser, local, nowPlaying] });

runtime.enqueue(NOW_PLAYING_URI);                         // the podcast already
                                                          // playing in their browser
runtime.enqueue('spotify:track:1OWBh1eVxUdA1Z6UA8r4nh');  // then a Spotify track
runtime.enqueue('https://example.com/episode.mp3');       // then a file on the web
runtime.enqueue('something calmer after those');          // then whatever you decide, later

await runtime.play();

Four sources, one list, in order. The first one is playing inside an app you do not own — and the queue waits for it to finish before taking over.

What you get

Add audio to your product without marrying one service. Write against one queue; swap or add backends later. An entry describes what to play, not where from, so it can bind to whichever source is available at the moment it plays.

Join what someone is already listening to instead of talking over it. The machine's current playback — a YouTube tab, a podcast in Safari, VLC — can be a queue entry like any other. Your track starts when theirs ends.

Know what you can do before you try it. runtime.can('seek') answers for the backend that is actually loaded. No silent no-ops, no discovering at 2am that one source quietly ignored a command.

Keep playing when a source fails. If a backend cannot load an entry, the same description is handed to the next one that can. A queue does not stop because one service is down.

Survive a restart. serialize() / restore() — and a queue saved on a machine with Spotify reopens on one without it, then plays from somewhere else.

What ships today

What each one can actually do

The point of the capability model is that these differ, and say so:

¹ only when you point it at a music folder to index · ² the AppleScript dictionary cannot search a catalogue · ³ there is no way to ask macOS's Now Playing register to start a specific track

Every ❌ there is a refusal rather than a silent failure. An adapter that claims it can seek and then doesn't is a bug you chase for an hour; these tell you first, and the runtime routes around them.

Not built yet: Now Playing on Windows, and controlling one specific browser tab (needs an extension). Details at the bottom.

---

The problem

Every audio integration today puts the queue in the wrong place.

   WITHOUT upnext                          WITH upnext

   caller                                  caller
     │                                       │
     │ "play X"                              │ enqueue / move / skip
     ▼                                       ▼
   Spotify Web API                     ┌───────────────┐
     │                                 │   THE QUEUE   │ ← yours. one of them.
     ▼                                 └───────┬───────┘
   ┌───────────────┐                           │
   │ Spotify queue │ ← the real one    ┌───────┼───────┬────────┐
   └───────────────┘                   ▼       ▼       ▼        ▼
                                    Spotify  Apple  browser   local
   Now queue a YouTube video.                 Music    tab      file
   Nowhere to put it.               Each one just plays what it's handed.

That works right up until the next item isn't a Spotify track — and then there is nowhere to put it. upnext inverts it: the runtime owns the queue, and Spotify's queue, Apple Music's Up Next and a browser tab's <audio> element all become places to send one item at a time.

---

Quickstart

The fast way — every source this machine can reach, one call:

npm i upnext-desktop

import { desktop } from 'upnext-desktop';

const audio = await desktop();
audio.enqueue('spotify:track:1OWBh1eVxUdA1Z6UA8r4nh');
audio.enqueue('https://example.com/podcast.mp3');
await audio.play();

It ships a CLI too:

$ upnext now
▶ Korea's STRANGEST Food is on Jeju Island!! — More Best Ever Food Review Show
  Google Chrome · 23:20 / 24:11

That is a YouTube tab, read with no browser extension.

Or give it to an agent

{ "mcpServers": { "upnext": { "command": "npx", "args": ["-y", "upnext-mcp"] } } }

Twelve tools in Claude Desktop, Cursor or anything else that speaks MCP — including media_adopt_current, which puts what someone is already listening to into the queue so the agent adds to it rather than talking over it.

Or wire it yourself

npm i upnext-core upnext-adapter-local

upnext-adapter-local needs ffplay (from ffmpeg) or afplay (built into macOS).

import { Runtime } from 'upnext-core';
import { LocalAdapter } from 'upnext-adapter-local';

const runtime = new Runtime({
  // Absolute paths only — Node does not expand `~`.
  adapters: [new LocalAdapter({ library: ['/Users/you/Music'] })],
});

runtime.on('item:started', ({ item }) => console.log('▶', item.ref.title));
runtime.on('item:ended',   ({ item }) => console.log('■', item.ref.title));

runtime.enqueue('file:///path/to/first.mp3');
runtime.enqueue('file:///path/to/second.mp3');

await runtime.play();     // plays the first, then the second, on its own
await runtime.next();     // skip

On a Mac with Spotify open, add a second source with nothing to sign up for:

npm i upnext-adapter-spotify

import { SpotifyDesktopAdapter } from 'upnext-adapter-spotify';

const runtime = new Runtime({
  adapters: [new LocalAdapter({ library: ['/Users/you/Music'] }), new SpotifyDesktopAdapter()],
});

runtime.enqueue('https://open.spotify.com/track/1OWBh1eVxUdA1Z6UA8r4nh');
runtime.enqueue('file:///path/to/second.mp3');
await runtime.play();   // Spotify, then a local file, without either knowing

Want to hear it right now, with no files of your own?

git clone https://github.com/tothienbao6a0/upnext && cd upnext
npm install && npm run demo    # synthesizes its own tones and plays them

---

How it works

The pieces

                          your application
                                │
   ┌────────────────────────────┼────────────────────────────┐
   │  Runtime                   ▼                            │
   │  ┌──────────┐  ┌──────────────┐  ┌────────────────┐     │
   │  │  Queue   │  │    Binder    │  │   Prefetcher   │     │
   │  │ ordered  │  │ which source │  │  resolve ahead │     │
   │  │ id-based │  │  + fallback  │  │  of the head   │     │
   │  └──────────┘  └──────────────┘  └────────────────┘     │
   │  ┌──────────┐  ┌──────────────┐  ┌────────────────┐     │
   │  │   Deck   │  │   Watcher    │  │   Reconciler   │     │
   │  │  loaded  │  │  is it over  │  │  human took    │     │
   │  │   item   │  │     yet?     │  │      over      │     │
   │  └──────────┘  └──────────────┘  └────────────────┘     │
   └────────────────────────────┬────────────────────────────┘
                                │  Adapter interface
    ┌──────────┬──────────┬────────┴───┬──────────┬──────────┐
    ▼          ▼          ▼            ▼          ▼          ▼
 Spotify    Spotify    a media      local     whatever    anything
 desktop    Web API     element     files      is on      you write
                     (browser /              (macOS Now
                      Electron)               Playing)

The life of one queue entry

  runtime.enqueue('something calmer')
              │
              ▼
        ┌──────────┐   your resolveIntent()    ┌────────────┐
        │ pending  │ ────────────────────────► │ unresolved │
        └──────────┘   "calmer" → a MediaRef   └─────┬──────┘
                                                     │ Binder picks a source
                                                     ▼
                                               ┌───────────┐
                                               │   ready   │  bound, not yet playing
                                               └─────┬─────┘
                                                     │ play()
                                                     ▼
                                               ┌───────────┐
                                               │  active   │  ◄── Watcher is watching
                                               └─────┬─────┘
                              ┌──────────────────────┼──────────────────────┐
                              ▼                      ▼                      ▼
                        ┌──────────┐           ┌──────────┐           ┌──────────┐
                        │  ended   │           │ skipped  │           │  failed  │
                        └──────────┘           └──────────┘           └──────────┘

Entries are prepared before the playhead reaches them, so an intent has already become a real track on a real backend by the time it's needed — no silence while a model thinks.

What play() actually does

1.  detach the deck, stop whatever was playing
2.  mark the entry loading
3.  intent?  ──► call your resolveIntent()          ──► MediaRef
4.  Binder scores every adapter with match(ref)
5.       ├─ resolve()      ──► a Binding
6.       ├─ confidence check: is this actually the song asked for?   ← ⚠ the big one
7.       ├─ load()
8.       └─ play()                    any step fails ──► try the next source
9.  Deck attaches, Watcher arms end-of-track detection
10. emit item:started + one queue:changed

---

The two ideas that make it work

1. Media is described, not located

A queue entry is not a URI. It's a MediaRef — a description that binds to a source as late as possible.

{ title: 'Bad Habit', artist: 'Steve Lacy', isrc: 'USUM72209293' }

Strong external ids (ISRC, MusicBrainz) are the join key; normalized title and artist are the fallback. Resolutions are verified before they play — an adapter returning something is not the same as it returning the right thing, and confidently playing the wrong song is the classic cross-source failure.

A title needs a backend that can search. A link says exactly what to play; a title has to be looked up, and not every backend can look things up. The Spotify desktop app is the sharp case — it plays a URI you hand it, but its AppleScript dictionary cannot search a catalogue, so it scores 0 for a bare title rather than guessing. On a default Mac setup that means nothing resolves { title: 'Bad Habit' }.

On a Mac this is now answered for you: upnext-adapter-apple-music searches your library and needs no credentials at all, so a plain title resolves out of the box. Elsewhere, index a music folder, add a Spotify Web token, or supply resolveIntent and answer it yourself. upnext-desktop's explainSetup() and upnext doctor both say which of those you have — this is a real gap and it is better named than discovered.

2. Capabilities describe what a backend actually is

Every backend sits somewhere on this line, and the runtime is correct across all of it:

  you own it completely  ◄──────────────────────────────►  someone else owns it

  local file            browser tab          Apple Music         Spotify app
  ───────────           ───────────          ───────────         ───────────
  process exit          'ended' event        must be polled      must be watched
  = end of track        = end of track
  exact position        exact position       exact position      exact position
  nobody else           nobody else          A HUMAN CAN         A HUMAN CAN
  can touch it          can touch it         HIT NEXT            HIT NEXT

{
  endOfTrack:      'event',      // 'event' | 'poll' | 'none'
  position:        'estimated',  // 'authoritative' | 'estimated' | 'none'
  externalControl: true,         // can a human change this behind our back?
  seek: true, pause: true, volume: false, search: true,
}

play: true would be useless — every adapter can play. These are the flags that change what the runtime and the caller actually do:

They're published inline on playback state, so this is one call, not a join against adapterId:

if (runtime.can('seek')) await runtime.seek(30_000);

Worked example: the same service, twice

upnext-adapter-spotify ships two adapters for Spotify, and they are not interchangeable:

search: false is the interesting one. Spotify's AppleScript dictionary cannot search a catalogue. That could be faked — scrape something, guess — and then every resolution of a title would be a coin flip dressed as a lookup. An adapter that says it cannot do a thing is correct and slightly limited; one that says it can and then does it badly is broken. So it declares false, scores 0 on anything that isn't already a Spotify link, and the entry goes to a backend that can actually find it.

That is the whole capability model in one flag, and it is why capabilities belong to an adapter rather than to a service.

---

When a human takes over

You queue three songs. The listener picks up their phone and hits next in Spotify.

   runtime thinks:   ▶ Nights          backend is actually playing:  ▶ Ivy
                       ↑                                               ↑
                       └───────────────── desync ──────────────────────┘

   policy 'adopt'    (default)  →  Ivy becomes a real queue entry, playback continues
   policy 'correct'             →  force the backend back to Nights
   policy 'ignore'              →  report it, change nothing

The human wins by default. A queue that fights the person holding the keyboard is a bug, not a feature.

The hard part is that "the track I loaded is not the track that is playing" has two opposite causes — our track ended and the backend rolled on, or a person chose something else — and they call for opposite responses. What separates them is where the playhead was a moment ago, which is knowledge only the adapter has. See adapter-spotify/src/sampler.ts for the real one.

---

Intents are queue entries

runtime.enqueue('something calmer after this');

That entry stays unresolved until the playhead gets close, then calls the resolver your host supplies:

new Runtime({
  resolveIntent: async (intent, ctx) => {
    // ctx.nowPlaying is what the listener actually just heard
    return await yourModel.pickTrack(intent, ctx);
  },
});

The core never calls a model, never holds an API key, never picks a provider. That boundary is what makes this embeddable in someone else's product instead of being one agent with a package.json. The Spotify adapter draws the same line around OAuth: you supply getAccessToken, it runs no flow.

Without a resolver it falls back to searching whatever adapters advertise search, so it's useful with nothing but adapters wired up.

---

API

// queue — always addressed by stable id, never by index
runtime.enqueue(input, position?)          // MediaRef | uri string | intent string
runtime.enqueueMany(inputs, position?)
runtime.move(id, { after: otherId }, expectVersion?)
runtime.remove(id, expectVersion?)
runtime.clear({ keepActive })

// transport
runtime.play(id?)        runtime.playNow(input)
runtime.pause()          runtime.resume()        runtime.toggle()
runtime.next()           runtime.previous()
runtime.seek(ms)         runtime.setVolume(0..1)  runtime.stop()

// how the queue is traversed
runtime.setRepeat('off' | 'one' | 'all')
runtime.setShuffle(true)

// surviving a restart
runtime.serialize()      // plain JSON — store it wherever
runtime.restore(state)   // → { positionMs }; replaces the queue, starts nothing

// reading
runtime.can(capability)  // what the loaded backend supports, right now
runtime.getState()       // { version, repeat, shuffle, playback, nowPlaying, queue, adapters }
runtime.queue            // frozen read-only view
runtime.search(query, { limit, adapterId })

// events
runtime.on(event, handler)  // returns an unsubscribe function

Repeat and shuffle live above the backends

Spotify has a repeat button. So does Apple Music. Neither knows about the browser tab queued behind it, so the only place the question can be answered once is above all of them. The adapters don't touch their backend's own setting.

Two details worth knowing: repeat-one still yields to next(), because a repeat mode that ignores the skip button is a trap; and shuffle is a traversal order, not a re-ordering — your list stays in the order you built it, and the runtime just picks differently. Inject random to make a shuffle reproducible in a test.

Surviving a restart

await fs.writeFile('queue.json', JSON.stringify(runtime.serialize()));

// …next launch
const { positionMs } = runtime.restore(JSON.parse(await fs.readFile('queue.json', 'utf8')));
await runtime.play();
await runtime.seek(positionMs);

Restoring never starts playback — that's the host's call. And bindings are dropped: a binding is a live handle to a backend session, and none of that survives a restart, so every entry rebinds against the adapters that exist now.

Which is the payoff for describing media instead of locating it: a queue saved on a machine with Spotify reopens on one without it, and still plays from somewhere else.

Why positions are ids, not indexes

   caller reads queue:  [0] Nights  [1] Ivy  [2] Pyramids
   caller decides:      "move index 2 to the front"
   meanwhile a human:   removes Nights
   the call lands:      moves Ivy.  Wrong song. No error.

So it's move(id, { after: otherId }). Every mutation bumps a version, and any mutation can pass expectVersion to refuse a stale write.

Events

Everything handed out is a copy, including event payloads. runtime.queue is a frozen view with no mutators on it — not a type-level Readonly a cast could defeat.

---

Failure is a first-class case

Three things look identical from a listener's chair — nothing is playing — so the runtime tells them apart:

---

Writing an adapter

Required: id, capabilities, match, resolve, load, play, stop. Everything else is optional and gated by what you declare — a thirty-line adapter is a legitimate adapter.

import { defaultCapabilities, type Adapter } from 'upnext-core';

class MyAdapter implements Adapter {
  id = 'mine';
  capabilities = { ...defaultCapabilities, endOfTrack: 'event', pause: true };

  match(ref)          { return ref.uri?.startsWith('mine:') ? 1 : 0; }
  async resolve(ref)  { return { adapterId: this.id, nativeUri: ref.uri, ref }; }
  async load(binding) { /* … */ }
  async play()        { /* … */ }
  async stop()        { /* … */ }
  subscribe(listener) { /* call listener({ type: 'ended' }) when a track finishes */ }
}

Two rules that matter more than the code:

1. Declare capabilities honestly. When in doubt, declare the weaker thing. A backend that says it can't seek is correct and slightly limited; one that says it can and then doesn't is broken.
2. Return null from resolve rather than guessing. The runtime tries the next source, which beats confidently playing the wrong song.

…in any language

Adapters don't have to be TypeScript, or even in this process.

   host                                 your child process
   ────                                 ──────────────────
   → {"id":1,"method":"init"}
                                        ← {"id":1,"result":{"capabilities":{…}}}
   → {"id":2,"method":"resolve", …}
                                        ← {"id":2,"result":{"nativeUri":"…"}}
   → {"id":3,"method":"play"}
                                        ← {"event":{"type":"ended"}}

One JSON object per line. No framing headers, no schema registry, no codegen. examples/python-adapter/adapter.py is a complete working backend in ~150 lines of Python, covered by the test suite — the runtime can't tell it apart from a native one.

---

Packages

upnext-core does no I/O at all — no filesystem, no network, no clock it wasn't handed. It runs identically in Node, Bun, Deno, Electron, Tauri or a browser, and the entire suite runs in milliseconds with no fake-timer library and no flakes.

---

Status

Early, but the core, the capability model and the adapter contract are real and tested. Every behavioural change carries a test, and CI runs the whole suite — plus the audible demo — on Node 20 and 22 across Linux and macOS on every push.

Several bugs in this design were found by running the demo out loud rather than by reading code — a doubled end-of-track event, a late prefetch overwriting the track that had just started, two tracks playing at once after a cancelled skip. Each has a regression test. If you touch playback, run npm run demo and listen.

Not built yet:

• The Windows equivalent of Now Playing. Windows has one — SMTC — and the adapter's shape would carry over, but nobody has written it. macOS (MediaRemote) and Linux (MPRIS, via playerctl) both work today, behind the same nowplaying:current entry.
• Controlling one specific browser tab. upnext-adapter-nowplaying already reaches whatever the machine is playing, browser included, through macOS's system Now Playing register — no extension needed. Singling out one tab among several, though, does need a browser extension, and that is a control feature rather than a queue one: you cannot queue into a tab you do not own.
• Gapless on backends we do not own. upnext-adapter-browser is gapless now — give it a spare element and the next track is buffered while the current one plays, so the switch is instant. Spotify and Apple Music are not, and cannot be from here: their AppleScript dictionaries have no way to hand them a track to play next, so the runtime has to drive each transition and that round trip is the gap. (All three transports ship now: upnext-mcp, the upnext CLI in upnext-desktop, and upnext-http.)

---

Contributing

The most valuable thing you can contribute is an adapter — see CONTRIBUTING.md. The core is deliberately small and mostly finished; what makes this useful is the number of places it can send audio.

Apache-2.0. Adoption is the only moat that matters for a substrate like this — a queue abstraction is worthless unless other people's adapters target it.