MusicGPT
AIGenerate songs, sound effects, lyrics, and speech, and reshape existing recordings, with the MusicGPT public API. Agents produce audio assets inside the workflow that needs them.
What This Integration Enables
MusicGPT is the widest connector in this part of the catalog, and it is easier to understand as an audio workbench an agent can drive than as a music generator. Yes, it writes full songs from a prompt, composes from an image, generates lyrics, and produces standalone sound effects and ambiences. That is the half people expect.
The half that changes what a workflow can do is what it does to recordings you already have. It separates stems, transcribes speech with word level timings and subtitle output, extracts key and tempo, masters a mix against a reference, converts formats, converts audio to MIDI, and strips noise, echo, and reverb. It also reshapes performances: remixing a track into a new style, extending it past a timestamp, replacing a singer's voice while keeping the timing, and rewriting a specific time range inside a recording. Nearly every operation is asynchronous and returns a task or conversion ID that Get Conversion polls, so a flow can fan out a dozen jobs and collect them as they land. That last group of capabilities is powerful enough that the human-in-the-loop gate on this connector should sit on the recordings, not on the generated music.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Podcast post-production as one flow
A raw episode lands in the folder. The agent runs Remove Noise and Remove Reverb to clean a take recorded in a hard room, transcribes the cleaned audio with Transcribe Audio to get both a readable transcript and SubRip subtitles, and cuts the two pull quotes the producer flagged with Cut Audio. It renders a short intro bed with Generate Music, writes the transcript and the show notes into Notion, and posts the clip URLs into Slack for the social team. What used to be four applications and three downloads becomes one run with a single approval at the end.
Briefed audio assets for a production queue
A game or video team files requests as written briefs. The agent drafts a tighter prompt with OpenAI, then routes it by asset type: Generate Sound Effect for a single ambience or effect, Generate Music for a full cue, Generate Song From Image when the brief is a mood board rather than a sentence, and Generate Lyrics when a writer needs a starting point. Each job returns a task ID; the agent polls Get Conversion and files the finished asset against the request that asked for it, so the requester gets a URL rather than a place in a queue.
Making a back catalog usable
An archive of recordings has no consistent loudness and no metadata worth searching. A scheduled flow walks the catalog, calls Extract Key And BPM to tag each track with its key and tempo, runs Extract Stems where an editor needs isolated vocals or instrumental, and passes each mix through Master Audio against one approved reference so the whole library sits at the same level. Convert Audio Format produces distribution copies at the target sample rate. The reference track itself is chosen by a person once, because it sets the sound of everything downstream.
Human-in-Loop Highlight
Inpaint Audio is the operation that earns the gate on this connector. It replaces a specific time range inside a recording with newly generated audio and blends it into the surrounding material, which is genuinely useful for fixing a bad take or a clipped word. Applied to a recording of a conversation that will be published, it does something else: it changes what the recording says happened, and it does it well enough that a listener will not notice.
So the agent files any inpaint on spoken material with the automation exceptions it is required to escalate, rather than treating it as an edit it may simply make. It stops, posts the passage to the producer with the original and the proposed range, and asks in Slack: "Episode 47, 12:31 to 12:38. The guest misstates the launch date here. I can trim at the sentence boundary and lose eight seconds, or rewrite the range with Inpaint Audio so the corrected date is spoken. Rewriting changes what the recording says the guest said. Which do you want, and does the guest know?" The agent holds until it gets an answer, and the answer is recorded against the episode. Trimming is an edit. Rewriting is a decision, and a person makes it.
Agent Capabilities
25 actionsMusic Generation
4- Generate Music Start an asynchronous full song generation from a natural language prompt, an optional style, and optional custom lyrics. Prompts under 280 characters give the most reliable results. Supports forcing an instrumental or vocal only render, applying a cloned voice, choosing a model version, and requesting an album cover. Returns a task ID plus up to two conversion IDs to poll with conversion type Music AI.
- Generate Song From Image Analyse an image and compose a song inspired by what it depicts, optionally steered by an extra prompt, custom lyrics, a musical key, and a target tempo. Negative tags exclude themes you do not want. Used when the brief arrives as a mood board rather than a sentence.
- Generate Lyrics Write original lyrics from a theme, vibe, or idea expressed in plain language. Reusing a prompt returns a variation rather than identical text, and the response carries the estimated credit cost. This is synchronous and produces no audio, which makes it a cheap first step before committing to a render.
- Generate Sound Effect Create a standalone sound effect or ambience from a text prompt, with a beta control over duration. Runs asynchronously; poll with conversion type Sound Generator. Used for game audio, video stingers, and podcast transitions that do not warrant a full cue.
Reworking Existing Recordings
5- Remix Audio Reimagine an existing track in a new style described by a prompt, with optional lyrics and a vocal gender steering any newly generated singing. Poll with conversion type Remix.
- Extend Audio Continue an existing track past a chosen timestamp by generating new audio matched to its style, guided by an optional prompt and lyrics for the new section. Used to turn a short hook into a full arrangement or to lengthen a bed to fit a video.
- Inpaint Audio Replace a specific time range inside a track with newly generated audio, blended into the surrounding material. Built for fixing a bad take, swapping a section, or censoring a passage without re-rendering the whole piece. On spoken recordings this changes the record of what was said, so route it behind an approval.
- Sing Over Instrumental Add an AI vocal performance of your lyrics on top of an existing instrumental, with the singing style set by a prompt. Used to finish a backing track you already own rather than generating a new song around it.
- Create Cover Song Produce a cover of an existing song performed by a different AI voice while keeping the arrangement intact, with pitch shiftable by up to twelve semitones to suit the new range.
Voice
4- Convert Text To Speech Synthesize spoken audio from text. A supplied voice sample URL takes priority, then a voice model ID, then the gender preference, so a clean sample of the target speaker gives the best likeness. Poll with conversion type Text To Speech.
- Change Voice Replace the singer or speaker in an existing recording with a chosen AI voice model while keeping the original performance and timing. Pitch shifts up to twelve semitones in either direction, and background noise can be stripped first.
- List Voices Return a page of the MusicGPT voice catalog with the voice IDs required by the generation, cover, and voice changer actions. Page size and page number walk the catalog.
- Search Voices Look up voices by name and return their IDs, which is the quickest way to resolve a human readable voice name into the identifier the generation actions expect.
Audio Analysis and Transcription
3- Transcribe Audio Transcribe spoken content with automatic language detection, optional translation into English, and word level timestamps. Output can be plain text, formatted text, SubRip subtitles, or WebVTT, which covers captioning as well as note taking.
- Extract Stems Separate a recording into individual stems such as vocals, instrumental, male vocal, and female vocal, optionally running denoise, de-echo, or de-reverb preprocessing first. Eco mode extracts the requested stems in parallel for faster turnaround.
- Extract Key And BPM Analyse a recording to report its dominant musical key, any key changes over time, and its tempo. Used to tag a library, match tracks, or confirm generated music fits an existing arrangement.
Audio Repair and Conversion
8- Remove Noise Strip background hiss, hum, and ambient noise while preserving the intended signal. Usually the first step before transcription or before mastering a field recording.
- Remove Echo Remove slap-back and repeated echo artefacts from a recording made in a reflective space, leaving a tighter and more intelligible signal.
- Remove Reverb Suppress room reverberation so a vocal or spoken take sounds close-miked, which is often required before voice conversion or stem separation.
- Master Audio Master a mix so its loudness, tonal balance, and dynamics match a reference track you supply, giving a consistent sound across a catalog. Source and reference are both given as URLs, and the output format is selectable.
- Cut Audio Trim a recording to the range between a start and end point in milliseconds, writing the excerpt out in the format you choose. Used to pull a hook out of a song or a quote out of an interview.
- Change Audio Speed Speed up or slow down a recording by a factor between 0.25 and 4.0. Used to fit narration to a fixed video duration or to produce slowed edits.
- Convert Audio Format Transcode a recording into a different container, sample rate, and bit depth, such as turning a 48 kHz WAV master into a 44.1 kHz MP3 for distribution.
- Convert Audio To MIDI Transcribe a recording into a MIDI file of the notes it detects, optionally rendering a sonified preview and a CSV of the individual note events. Turns a hummed idea or a live take into editable notation for a DAW.
Job Status
1- Get Conversion Retrieve the state and results of any MusicGPT job using either the task ID or a conversion ID returned when the job was submitted, with the conversion type matching the action that created it. Once processing completes the response carries the hosted audio URL and related metadata, which makes this the polling companion to every asynchronous action here.
Frequently Asked Questions
What can FlowRunner do with MusicGPT?
FlowRunner agents can run Generate Music, Generate Sound Effect, and Generate Lyrics in MusicGPT, plus 22 more actions.
Does connecting MusicGPT to FlowRunner require OAuth?
No. MusicGPT connects to FlowRunner with an API key, no OAuth flow required.
Can MusicGPT trigger a FlowRunner workflow automatically?
MusicGPT doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with MusicGPT
$100 in credits. No card required. Connect in minutes.