ace-step-studio

ACE-Step Studio through MCP

ACE-Step Studio serves MCP at http://127.0.0.1:8792/mcp while it is open (Streamable HTTP, stateless JSON-RPC). Every tool runs the same code as a button of the studio, and the user sees what you do in the studio's window.

If the studio is not running yet

1. It is a Windows desktop application. If it is not installed, download the installer or the portable archive from https://github.com/timoncool/ACE-Step-Studio/releases/latest (it needs an NVIDIA card; the first start offers to download the models).
2. Start it. The MCP server is up as soon as its window is: http://127.0.0.1:8792/mcp. Nothing else to install - no npx, no bridge.
3. Connect (below), then call studio_status. If the models are missing, models_catalog and models_download fetch them.

Connect

claude mcp add --transport http ace-step-studio http://127.0.0.1:8792/mcp

Other clients: { "mcpServers": { "ace-step-studio": { "type": "streamable-http", "url": "http://127.0.0.1:8792/mcp" } } }.

The server also serves this skill (resource studio://skill, prompt studio) and the writing guides (resources studio://guide/<topic>). It speaks MCP 2026-07-28 (stateless: every request carries its version in _meta, server/discover describes the server) and the handshake revisions 2025-11-25, 2025-06-18 and 2025-03-26 through initialize. Only this computer's agents and the studio's own window may connect.

The user sees it in the studio too: Settings, Agent (MCP) shows whether an agent is connected and the address to paste.

Ground rules

• Start with studio_status. It tells what runs now and whether the window is open.
• Long work is a job: songs, stems, MIDI, karaoke, preparation, training. Start it, then studio_wait (a job_id, or until: stems | midi | processing | covers_and_karaoke | song_jobs | preparation | training | idle) instead of polling. It returns within a minute (30 s by default, 55 at most) with how far the work got; call it again.
• One heavy job holds the graphics card at a time. While a LoRA trains no song is made; start training last.
• Answers are short by default: a song job is its status and the songs it made, a library song leaves out its audio codes, lora_list gives one line per LoRA. Pass response_format: detailed when you need every field.
• Covers are drawn only with an image model set up (settings_get, covers; an OpenRouter key). Without one a cover_prompt is kept but no cover appears; cover_set_from_file still works.
• Look ids up, never guess them: library_songs_list, training_status, dataset_get, lora_list, models_status.
• What you change is written down where the user reads it. Every change you make - a generation started, a song edited or deleted, a playlist, stems, karaoke, a dataset and its preparation, training, a LoRA, a model downloaded or removed, a setting - goes into the studio's Activity journal as what was done, to what and its name; a job is written as started. Reads, the window's own controls (ui_*, the player, the create form, the video editor, the visualizer) and answers that store nothing leave no line. A tool that writes over what was stored (a song or playlist edit, a dataset prepared again, stems made again, karaoke, MIDI, a cover) is marked destructive, so your client asks the user first.
• A like lives with the song: library_liked (the latest like first) and library_song_like work without the window open.
• A stem belongs to its song: the window shows it under the song it was separated from, stems_split on a song that has stems replaces the ones it makes again, and a stem itself is not split.
• Files on this computer are passed by path: dataset_add_folder, library_import_audio, cover_set_from_file, video_set. library_song_files and dataset_song_files give the paths of the studio's own files.
• One writer per song. Either you write the caption and lyrics, or ACE-Step's own language model does (think on with empty lyrics, or song_plan) - never both on one song. You write better: read writing_guide and writing_examples first. Use assistant_write only when the user asks for the studio's assistant. assistant_sections tags lyrics the user wrote with their sections without changing a word - the tag button of the create form.