Skip to main content
G

dsh-ace-music

gladess1102/dsh-ace-music

ACE Music MCP server for DSH: sign in once with your own account over OAuth, then write full songs, continue a melody from a MusicXML or ABC score, upload audio for transcription, and inspect or download tracks as native acemusic_* tools. Ships a dependency-free MusicXML-to-ABC converter and ACE Music's published writing guides for offline use.

Install

dsh plugin --profile web add github:gladess1102/dsh-ace-music

README

dsh-ace-music

中文说明 · DeepSeek Harness plugin

ACE Music's MCP server inside DeepSeek Harness. Sign in once, then write whole songs — vocals included — and inspect, list or download them without leaving the conversation.


First, what ACE Music actually is

ACE Music v1 is ranked #1 on Artificial Analysis' AA-Music-Vocal v1.1 leaderboard with 1160 Elo — ahead of Suno v6 (1144), Suno v6-mini (1132), Mureka V9.5 (1101) and Lyria 3.5 (1050). That leaderboard is a blind listening test: evaluators hear tracks made from the same prompt without knowing which model produced them. It ranks vocal music specifically, not every genre feature — but vocals are exactly where generated music usually gives itself away.

Two models, two different jobs.

ACE Music v1 — words in, song out. Give it a style description, your own lyrics, or both. It is built for people who don't write music: describe "warm indie folk with acoustic guitar and a touch of nostalgia" and a finished song comes back. It also reads performance directions in parentheses — (hum softly), (sing with a raspy voice), (add guitar here) — so you steer not only what is sung but how the vocal and the arrangement develop.

ACE Music v1.1 Preview — a score you control. It plans melody and harmony as ABC notation before any audio exists. That score fixes the melody, rhythm and chords, and the song is then performed from it: compose first, perform second. Start from a melody you hummed, played, or wrote as MIDI; edit notes and chords; regenerate. The notation is what stays constant. As the model page puts it, each generation is a new performance, so the sound can differ even when the score does not.

Around those two: keyboard-played melodies, humming-to-song, image-to-song, AI cover, mashup, and stem separation, plus MIDI/score export of the main melody and chords.

Published pricing (USD): Free — 20 credits/day (about 10 songs), 10 upload minutes, shared queue, no downloads, no commercial rights. Creator $7.90/mo — 1,000 credits (about 500 songs), 30 upload minutes, unlimited downloads, priority queue, stem splitter, commercial rights. Artist $16.58/mo — 2,500 credits, 60 upload minutes, plus ACE Studio. One generation costs 4 credits and returns two songs. The open ACE-Step 1.5 model is free to run yourself, with no limits.

Live numbers, not marketing copy: acemusic_get_account reports your actual balance.

Before you install: create an ACE Music account

The plugin signs in as you, with your own account — it has no key of its own and cannot create accounts.

  1. Open https://acemusic.ai and register (email or the sign-in providers it offers).
  2. Stay signed in in that browser — the plugin opens an authorization page there, and being signed in makes it one click.
  3. Optional but useful: generate one song on the website first. You will learn what the style and lyrics fields do, which is exactly what acemusic_create_song sends.

Signing in costs nothing. A free account can do everything except download the audio and use results commercially.

Install

# from npm
dsh plugin --profile <your-profile> add dsh-ace-music

# or straight from GitHub
dsh plugin --profile <your-profile> add github:Gladess1102/dsh-ace-music

Then in DSH: Plugins → Add plugin → point at the installed package, or restart — the bundle patch activates on load.

Sign in

acemusic_login

A browser tab opens on acemusic.ai showing "Connect DeepSeek Harness (ACE Music plugin) to ACE Music". Press Allow. The token lands in ~/.dsh/ace-music/credentials.json (mode 0600, your user only) and every acemusic_* tool becomes available.

acemusic_status

reports the plan, remaining credits, refill time and upload minutes. Tokens last 3600 seconds and refresh automatically; acemusic_logout deletes the stored token.

Tools

Ten tools are bridged from the remote server and registered as acemusic_*:

ToolWhat it does
acemusic_create_songStyle (required) + lyrics + title → a song. Optional ABC/MIDI score on v1.1.
acemusic_continue_scoreContinue a melody from an ABC or MIDI opening (v1.1).
acemusic_upload_audioAdd a recording or demo to your library; it is transcribed.
acemusic_get_skillACE Music's own writing guides (cover, lyrics, mashup, score-format).
acemusic_get_songStatus (writing → rendering → ready / failed), title, style, lyrics, page link.
acemusic_list_songsPage through your songs and uploads, newest first.
acemusic_download_songDownload link (MP3 / WAV / MIDI). Paid plans only.
acemusic_show_songsPlayer card — meaningful in Claude/ChatGPT, plain text in DSH.
acemusic_show_upgradeUpgrade card — plain text in DSH.
acemusic_get_accountPlan, credits, refill schedule, upload minutes, available models.

Plus six local tools:

ToolWhat it does
acemusic_loginStart the OAuth flow; returns the authorization link and waits in the background.
acemusic_statusSign-in state + live account info.
acemusic_logoutDelete the stored token.
acemusic_toolsRe-fetch the remote tool list and register it (idempotent).
acemusic_callEscape hatch: call any remote tool by name with raw JSON arguments.
acemusic_scoreConvert MusicXML into ACE's ABC dialect, ready for create_song's score.abc.

Bundled writing guides

The four guides ACE Music publishes through get_skill ship inside this package under docs/skills/:

FileRead it before
lyrics.mdWriting any song — section tags, parentheses cues, line counts, length
cover.mdMaking a cover — how to hand over score and lyrics for another style, key, tempo or language
mashup.mdJoining two or three songs — one key and tempo, sections taking turns
score-format.mdWriting or editing ABC notation, and how MIDI is read

acemusic_get_skill returns the local copy first, so the model can consult the rules with no network round trip; the reply says so, and points at acemusic_call if you want the live version. Their text belongs to ACE Music (© Timedomain, Inc.) and is cached here verbatim — only this project's code is MIT.

MusicXML as the writing layer

ACE Music's MCP endpoint takes a score as ABC or MIDI only — yet its web app accepts .musicxml, .xml and .mxl (Idea → Sing sheet music, on ACE Music v1.1 Preview). That path is closed to an agent, which has no file-upload equivalent over MCP.

So the split this plugin settles on: write in MusicXML, deliver in ABC.

MusicXML is the better format for a model to compose in. Pitch lives in <pitch>, duration in <duration>, harmony in <harmony>, words in <lyric>; a bar is a <measure> you can add one at a time, and a bar whose durations do not add up to the meter is detectable by code. Compared with ABC, you are far less likely to write a chord as a note, or lose a beat without noticing.

acemusic_score performs the conversion:

MusicXMLABC
part with <lyric>V: Vocal
other partV: Ins (ACE accepts only these two lines)
<note> with <chord/>top note only
<harmony>"Dm" / "F/A" chord symbols
<direction><rehearsal>% verse, % chorus …
divisions + durationACE's L:1/16 units, falling back to 1/32
alter^ sharp, _ flat

Two things bite. Sections must match exactly: the count and order of % <section> markers must equal the [section] tags in the lyrics, or ACE answers score_invalid and charges nothing. Only two voices exist: a layered choir cannot be notated as four parts — write the top line, and describe the harmony in the style. The full contract and the traps are in docs/musicxml.md.

Measured on the reference piece — 46 bars, 72 BPM, D minor, female choir, acemusic-v1.1:

Sectionlow band <200 HzRMS
intro6.4%−24.3 dB
verse4.7%−20.9 dB
chorus6.9%−17.5 dB
bridge (marked a cappella)1.4%−22.0 dB
final chorus7.9%−17.0 dB
outro1.5%−29.8 dB

Section structure, the dynamic arc and the a cappella drop all survived the trip from MusicXML to audio; the printed length (153 s) matched the render (157–160 s).

Why this is a plugin at all

https://acemusic.ai/mcp is a Streamable HTTP MCP server that accepts OAuth only — an unauthenticated call answers 401 with WWW-Authenticate: Bearer resource_metadata=....

DSH's built-in @deepseek-ai/dsh-mcp-client supplies static request headers and nothing else (createTransport never passes an authProvider), so it cannot answer that challenge. Servers that accept a personal access token — GitHub and friends — work fine, which is why the gap is easy to miss.

This plugin closes it: dynamic client registration, PKCE authorization code flow, a loopback callback on 127.0.0.1, token exchange and refresh, then the Bearer token bridges the remote tools into native DSH tools. No MCP SDK, no runtime dependencies — a small JSON-RPC client over fetch.

Configuration

VariableDefaultMeaning
ACE_MUSIC_PORT8765Loopback port for the OAuth callback. Change it if the port is taken.

State lives in ~/.dsh/ace-music/credentials.json (0600). Delete the file — or call acemusic_logout — to unlink the account.

Privacy and security

  • Your password is typed only on acemusic.ai; the plugin never sees it. It receives an OAuth token you explicitly granted, scoped to songs:read songs:create.
  • The token is stored outside the DSH configuration, so it never appears in cordis.patch.yml, logs, or a copied profile.
  • Network calls go to acemusic.ai and nowhere else. No telemetry, no third-party proxy.
  • No model weights are downloaded or run locally; generation happens on ACE Music's servers, under your account and your quota.

If the system resolver fails on acemusic.ai (some VPN and split-DNS setups do), the client falls back once to resolving through public DNS — the normal path is a plain fetch.

Limits worth knowing

  • Generation is slow. A song takes tens of seconds to a few minutes. acemusic_create_song has a 5-minute tool timeout; a timeout is not a failure — poll acemusic_get_song with the id.
  • Every generation spends credits. 4 credits per request, which returns two songs.
  • The free plan cannot download. acemusic_download_song answers with an upgrade prompt; playback on the song page stays free. Free-tier songs also carry no commercial-use rights (upgrading later does not grant them retroactively), and generation runs in a shared queue that can be busy.
  • show_songs / show_upgrade are designed for clients with card renderers; in DSH they return plain text.
  • The remote tool list is fetched at startup when a token exists and can be re-fetched with acemusic_tools; if ACE Music adds tools, that call picks them up.

Development

npm test         # 8 cases over the MusicXML converter
npm run check    # syntax check every module
FileRole
index.jsPlugin entry: apply(ctx), tool registration, remote bridging, local-first guides
src/oauth.jsDynamic client registration, PKCE, callback server, token exchange/refresh
src/mcp.jsMinimal MCP client (initialize / tools/list / tools/call, JSON and SSE)
src/net.jsfetch wrapper with timeout, retry, and a DNS fallback path
cordis.patch.ymlBundle patch: one plugin row, id: ace-music
docs/skills/ACE Music's published writing guides, cached for offline use

Protocol details are in docs/DETAIL.md.

License

MIT for the code. The bundled guides under docs/skills/ are ACE Music's content, cached with attribution. This project is not affiliated with ACE Studio or ACE Music; it is an independent client for their public MCP endpoint.

Related plugins