The studio

The studio turns clips into one piece of audio. It has a timeline with lanes, a mixer, a device rack and an export. It works on a story: a saved document whose clips have a place in time and keep it when they are regenerated.

There are two ways in:

• Voice Studio → new story. Write a script. A TTS model reads each line and the lines land on the timeline in order.
• Any sound → "Open in the studio". The clip is placed on a lane. If a separator is installed, pick the stems you want and each one gets its own lane.

Both make the same kind of document. You can add a spoken line to a story that started as stems, or drag a music bed under a script.

The layout

Three columns. Drag the divider between two columns to resize them.

Column	What lives there
Left (⅕)	The Script panel, the Prompt panel, or both
Middle	The Timeline, with the Device rack beneath it
Right (⅕)	The Library, the Sample and the History panels

The timeline has no close button. Close every other panel and it takes the whole page. A column with no open panels is not rendered, so closing the script leaves no gap; the timeline moves left into the space.

Below roughly tablet width the columns stack in order: left panels, timeline, right panels. No panel is hidden.

Panels in a column share the remaining height by weight. Drag the divider between two open panels to trade height between them, or focus it and use the arrow keys. The chevron collapses a panel's body but keeps it mounted. The timeline owns the audio graph, so collapsing it does not stop playback.

The panels

Every panel is available in every story. Which ones start open depends on the kind of story. This is decided once, when its clips first load, and is not changed afterwards.

Panel	Shortcut	What it is for	Open by default
Timeline	—	The arrangement. Always open, no close button	Always
Script	⌘/Ctrl + R	The lines, their voices and their delivery; write, edit, re-read	Yes, unless the story is an arrangement
Prompt	⌘/Ctrl + P	Describe a sound or a piece of music and drop the result onto this timeline	Only on a story with no clips yet
Library	⌘/Ctrl + B	Your sounds, to drag onto a lane	Only on an arrangement
Sample	⌘/Ctrl + E	The selected clip: waveform, stems, and what generated it	No; double-clicking a clip opens it
Devices	⌘/Ctrl + F	One track's insert chain	No
History	⌘/Ctrl + I	Every prompt you have run into this timeline, kept across reloads	No
MIDI	⌘/Ctrl + N	Play or import notes, and render them onto a lane	No
Piano roll	⌘/Ctrl + O	The same notes, as something you edit	No
Logs	⌘/Ctrl + G	What the server was doing while this was made	No

You can open and close panels from three places, which always list the same panels:

• the Panels menu in the header, which shows each shortcut beside its name. It is the only way to reopen a panel you closed with its X;
• the shortcuts above;
• the command palette, ⌘/Ctrl + K. Type to filter by name, move with the arrow keys and press Enter to run. It lists panels only.

Script and Prompt can be open in the left column at the same time, so you can narrate over a generated track or generate a bed under a script.

An "arrangement" is not a stored mode. A story counts as an arrangement when every clip has audio and none has words. So music and SFX stories open with no script and the library already open, and writing a line from the timeline's right-click menu brings the script back.

Panel shortcuts avoid the letters the timeline already uses (A C D M S V X Z, and Y for redo). Both sets of shortcuts listen on window and neither stops the other, so a shared letter would fire twice. That is why Script, History and Devices use R, I and F instead of S, Y and D.

Working with clips

A clip is a segment on a lane: a start time, a length, and a window into the sound behind it. Nothing you do on a timeline modifies the underlying audio file.

Selecting

Gesture	What it does
Click a clip	Select just that one
Shift-click, or ⌘/Ctrl-click	Add or remove one clip, keeping the rest
Drag across empty lane	Rubber band. Shift or ⌘/Ctrl adds to the existing selection
⌘/Ctrl + A	Select every clip
Escape	Clear the selection, and disarm the scissors

One clip in the selection is the primary. The Sample panel describes it and the device rack shows its track. If you shift-click the primary out of the selection, another selected clip becomes primary, so those panels stay filled while any clip is selected.

The rubber band selects every clip it overlaps, including clips it only partly covers. This lets you catch a long clip by crossing it.

A press on empty lane starts as a playhead scrub. It becomes a rubber band only after the pointer moves about four pixels. Clicking empty track moves the playhead and does not clear the selection. Use Escape to clear it. Escape is handled by the lane area, not the window, so click in the timeline first.

Moving

Drag a clip to move it in time and between lanes. Drag any clip in a multi-selection and the whole selection moves together, keeping its spacing and lane offsets. If part of the group would go past the start of the story or past the last lane, the move is limited for the whole group, so the clips keep their positions relative to each other and do not pile up at the edge.

Pressing on a clip that is already in a multi-selection keeps the selection, in case you are starting a group drag. If you release without moving, only that clip stays selected.

A group drag is one undo entry.

The script reads in timeline order. Moving a clip re-orders the script to match, in the same frame the clip lands. The timeline order always wins.

Snapping

Dragged edges snap to other clips' starts and ends, the playhead, time zero, and the two nearest ruler ticks. The tolerance is 8 pixels, so it is the same at every zoom level. Hold Alt to turn snapping off.

Both ends of a dragged clip look for a target and the closer one wins, so you can butt a clip against the clip before it or the clip after it with the same gesture. A clip never snaps to itself or to other clips moving with it.

Trimming

Drag either end of a clip. Clips narrower than about 24 pixels have no trim handles, so you can always grab the middle to drag.

Trimming changes the window into the source; it does not cut the file. Dragging the start moves both where the clip begins in time and where it begins in the source, by the same amount, so the remaining audio stays in place. Dragging the end is limited by how much source is left after the in-point. A pause has no source, so it stretches freely. Lengthen a shortened clip to get the trimmed audio back.

Splitting

Three ways, all the same operation:

• Select a clip with the playhead inside it and press S, or click the scissors button.
• Click the scissors with the playhead outside the clip to arm them, then click a generated clip where you want it cut. Escape cancels.
• Right-click a clip → Split at playhead, or right-click a lane → Split all at playhead, which cuts every clip the playhead crosses and reports how many.

Only a clip that has audio can be split.

Fades and crossfades

Every clip wider than about 36 pixels has a fade handle in each top corner, set inside the trim strips so the two never overlap. Drag a handle inward to lengthen the fade. The fade shape is drawn over the waveform so you can see how far into the clip it reaches. Each fade is capped at half the clip, so the two fades never cross.

Overlapping clips on one lane crossfade automatically across the overlap, whether or not you set a fade. The crossfade is capped at the length of the shorter clip, so a short clip dropped on a long one crossfades over its own length. Without this, the overlap would sum to a bump in level and could click at both edges.

Fades and crossfades use equal-power curves, which keep the level flat through a crossfade (two summed linear ramps dip by 3 dB in the middle). The client and the export use the same rule, so the mixdown matches what you hear while arranging.

Duplicating, deleting, copy and paste

⌘/Ctrl + D duplicates and Delete (or Backspace) deletes everything selected. The toolbar's duplicate, mute and delete buttons act on the primary clip only.

⌘/Ctrl + C copies the primary clip. ⌘/Ctrl + X copies the primary and deletes the whole selection. The clipboard holds one clip, because paste lands at the single playhead. ⌘/Ctrl + V pastes at the playhead, on the copied clip's own track. A copied line carries its words, speaker and voice, so pasting a written line that has not been voiced yet keeps its text and voice settings.

Right-click a lane → Paste here pastes where you clicked.

Muting and looping

M mutes or unmutes the primary clip. Muted clips are left out of the export.

Shift-drag the ruler to mark a loop region. It snaps like everything else, and Alt turns that off too. With no region marked, the loop button loops the selected clip. Switching the loop off keeps the region: it is drawn solid while looping is on and dashed while it is off, so you can play the rest of the story and come back to the same marks.

Zoom and the playhead

The mouse wheel zooms around the pointer, so the point under the cursor stays put. Shift + wheel scrolls horizontally (the browser's own scroll). The toolbar's zoom buttons step by 1.5× between 8 and 400 pixels per second. Click or drag the ruler, or empty lane, to move the playhead.

Right-click

On a clip: Read this line / Regenerate (only for a written line), Split at playhead, Split into stems… (when a separator model is installed), Duplicate, Extend this track and Cover in a new style (when the clip's model supports them), Repaint, Mute / Unmute, Delete.

Repaint regenerates part of a clip's audio:

1. A dialog names the time window and asks what should be there instead (leave it empty to reroll with the clip's own description), a strength and a mode.
2. The window is the span marked on the ruler (shift-drag it), cut down to the clip you right-clicked. The menu item shows the seconds, for example Repaint 1.8s–5.3s. With nothing marked it is the whole clip.
3. The result goes back in the same place. The clip does not move and is not re-trimmed, and the audio outside the window is unchanged.

Undo restores the previous take. Both takes stay in the library either way. Repaint needs a model that can repaint; ACE-Step is the only one today. The waveform editor's toolbar has the same action, where you drag out the selection by hand for a finer range.

On empty lane the menu is headed with the time and track you clicked: Create voice line, New recording, New prompt (opens the Prompt panel and puts the cursor in it), Insert sample (opens the Library), Split all at playhead, Paste here, Add track.

Tracks

Each lane has a head pinned to the left edge of the scroll area, so it stays visible when you scroll into the middle of the piece.

Control	What it does
The number, in colour	The lane's index and its hue. Clips, head and rack all share it
The name	Click to rename. Enter commits, Escape cancels; blank clears it back to "Track n"
M	Mute this track
S	Solo this track
The slider	Track gain, −30 to +6 dB in 1 dB steps
The bin (on hover)	Remove the track

Solo overrides mute. Once any track is soloed, only soloed tracks are audible, even if they are also muted. A lane you will not hear is drawn at half opacity.

There are eight lane colours spread around the hue circle. Hues near the app's teal accent are left out so a lane never looks selected. Each new lane steps half way round the circle, so lanes 0 and 1 are opposite colours. A ninth lane repeats the first colour.

Adding a track: use the Add track row under the last lane, the lane right-click menu, or drag a clip down onto the empty lane at the bottom.

Removing a track removes its clips too (the same rule as Audacity). You are asked to confirm when the lane has clips, and not when it is empty. Undo does not cover it.

New audio always lands on a new track, below all existing tracks. This applies to a generated clip, a recording, and a drop onto a lane that already has something there. Clips that share a lane and a time crossfade, so new audio is kept off existing lanes to avoid changing your arrangement without you asking. Drag it onto another lane if you want it there.

Dropping a sound from the Library lands it on the lane you aimed at, starting where you let go. If something already occupies that span, it goes to the next free lane.

Track name, mute, solo and gain are stored with the story, and the export applies all four. Clip gain is stored too, but there is no timeline control for it; set it over the API or from an agent.

Recording

The record button captures from your microphone onto a new track at the playhead. The playhead position is taken when you press record, so moving the playhead during the take does not move it.

While recording you see an elapsed counter and a live input level meter. Check the meter to confirm the microphone is picking you up.

The lane right-click menu has the same action as New recording. The take still goes onto a new lane of its own.

The browser picks the container (WebM/Opus on Chrome and Firefox, MP4/AAC on Safari). Soundbox decodes it, converts it to WAV, uploads it to your library as Take HH:MM, and then places the clip. Takes must be library sounds because the timeline references clips by id and the export reads them from disk.

If permission is denied, no microphone is found or the browser is unsupported, the record button's tooltip says so. There is no dialog.

MIDI

Experimental. The panel works end to end, but a placed take is audio, not notes. See "What this costs" below before relying on it.

The MIDI panel (Ctrl/⌘ + N) sits under the timeline beside the rack. Play a line, then press Place on timeline to render it and drop it on a new track at the playhead, the same way a microphone take is placed.

There are three ways to play:

Way	When
A MIDI keyboard	Where the browser has Web MIDI. Chromium does; the macOS desktop build's WKWebView does not
The on-screen keyboard	Always. Click the keys, or focus it and use your computer keyboard
Import a .mid	To bring in an existing piece of music without playing it

Knobs, pads and the pedal

Use Mapping, at the right of the keyboard row, to make a controller's knobs, pads, wheels and pedal do something. Unmapped controls are ignored.

It has three parts:

• The monitor. A live readout of the last message: port, message type (CC 74, Note 36, pitch bend), channel and value. Check it first when a control does nothing. It shows whether anything is arriving, which channel it is on, and whether the control is bound.
• A layout. A starting mapping for a whole controller, listed below.
• Learn. Press it on any row, then move the control you want to bind. Notes are not sounded while Learn is armed, so setting up a mapping does not add stray notes to your take.

Layout	What it assumes
General MIDI	Only the three controllers the spec fixes: mod wheel (CC 1), volume (CC 7), damper pedal (CC 64). Works on any board
M-Vave SMK-25 / 25 II	The above, plus the usual 8-pad default of notes 36-43 and a CC 21-24 knob bank
Sustain pedal only	Nothing but CC 64
Nothing mapped	Keys and wheels; Learn the rest

Apart from the pedal and the mod wheel, the presets are common defaults and may not match your board. The M-Vave's pads and knobs can be reassigned in its own editor app. Check the monitor, and use Learn on any control that does not respond. Each control drives one target: binding a control that already drives something moves the binding to the new target.

Eleven things can be bound: Sustain (hold), Record, Preview, Clear, Quantise, Octave ±, Place dry and Generate instrument (momentary), and Cover strength and Roll tempo (continuous). Sustain works like a piano pedal: a key released while it is down keeps ringing and keeps its timing, so a pedalled phrase records as held notes.

The mapping is stored per browser, not per story, because it describes your hardware. It applies to every story and survives a reload.

Blocked, not denied. If Chrome has a remembered block for the site, requestMIDIAccess is rejected immediately without a prompt, so the panel has no Retry button. Allow MIDI devices from the padlock menu in the address bar and reload. That is the only fix.

Notes are held while you play and drawn on a strip as they arrive. Quantise snaps note starts to a grid in seconds (a take played without a click has no bar lines). It is a button that edits the take, so the roll always shows what will be rendered. It moves starts only and leaves note ends where you played them.

The piano roll

The Piano roll panel (Ctrl/⌘ + O) shows the same take for editing: a keyboard down the left edge, pitch on the vertical axis, bars across the top.

Gesture	What it does
Drag on empty grid	Draw a note; keep dragging to set its length
Drag a note	Move it in time and pitch
Drag its right edge	Change its length
Arrow keys	Nudge the selected note; Shift for an octave or a bar
Delete, or right-click	Remove it
Hold Alt	Defeat snapping, as on the timeline

Editing a note plays it, so you hear the pitch while you drag.

The MIDI panel and the piano roll edit one take. Notes you play in the MIDI panel appear in the roll, and edits in the roll are what gets rendered. The roll's grid is in bars and beats and uses the tempo; the MIDI panel's recorder works in seconds.

The instrument is a second pass

No model in the catalogue takes notes as input. The panel renders the notes with a plain synth, then a second pass applies the timbre you describe:

notes ──▶ [synth] ──▶ dry clip ──▶ [Cover, ACE-Step] ──▶ "solo violin, large hall"

The synth gets the pitch and timing exactly right, and the model is only asked for timbre over audio whose notes are already correct. The cover is rendered at the source's own length, so it lines up with the take.

Generate instrument does both steps in one press. It places the dry take and submits a cover of it. The performance lands on one lane and the cover on the next, so you can solo between them and re-cover the dry take if a prompt does not work.

The cover clip is trimmed to the take's length. ACE-Step will not render shorter than about five seconds, so a three-second phrase drawn in the roll comes back at 5.12s whatever length is requested. Longer takes track closely: 7.70s covers to 7.68s. The extra is silence at the end, because the material sits at the start of the render. The audio past the trim is still on disk; drag the clip's right edge to get it back.

• Keep the notes is the cover strength. High keeps your synth line and changes its sound; low lets the model write its own notes.
• Place dry skips the model and places only the synth render.

These controls are at the bottom of both the MIDI panel and the piano roll, so you can finish a take from either one: play it in and place it without opening the roll, or draw it in the roll without touching the keyboard.

The four Instrument presets are bright and harmonically rich, because the cover pass works better from a filtered sawtooth than from a sine. The render is normalised to -3 dBFS, so the level the model receives does not depend on how many notes overlap.

For room sound, use the Reverb device instead of the prompt. Asking the model for a room smears the note attacks; the rack's reverb can be turned down.

What this costs: the story keeps the WAV, not the notes, so a placed take cannot be edited as notes afterwards. Play it again instead. Clips are audio everywhere else in the studio, and storing notes in a clip would need a change to the story document.

The device rack

The rack sits under the timeline and holds one track's inserts. Pick the lane with the buttons in its header. Selecting a clip also points the rack at that clip's track.

Devices are inserts, so they sit before the lane's fader; a gate hears the take, not your monitoring level. A new device is added to the end of the chain. The Add device menu lists devices roughly in the order they usually go in a chain, so adding them top to bottom gives a sensible order. There is no drag-to-reorder: the chain runs in the order you added the devices. Each device can be bypassed without losing its settings.

Drag a dial up and down to change it (hold Shift for a four-times finer sweep). Dials also work from the keyboard: arrows nudge by one step, Shift + arrow by ten, Home and End jump to the ends. Wet/dry is a horizontal slider.

Devices are heard on the timeline; they are not baked into the export. The mixdown applies clip gain, track gain, mute, solo, trims and fades, but no devices. Applying the rack on export would need a second implementation of all eight devices in the server's DSP. The rack shows this warning where you add a device.

For the same reason, racks are stored in your browser, per story, and not in the story document. They do not travel to another machine, and clearing site data clears them.

The devices

Values are shown in the unit you would tune in, which is why several ranges below are in milliseconds.

Device	Parameter	Range	Default
Gate: shuts the track between phrases, so room tone does not build up	Thresh	−80 to 0 dB	−45 dB
	Smooth	10 to 500 ms	60 ms
Autotune: pulls a sung take onto the nearest note	Key	C … B	C
	Scale	Chromatic, Major, Minor	Chromatic
	Retune	0 to 400 ms	80 ms
	Amount	0 to 1	1
EQ: three bands (low, mid, high)	Low	−24 to +12 dB	0 dB
	Mid	−24 to +12 dB	0 dB
	High	−24 to +12 dB	0 dB
	Lo × Mid	80 to 1000 Hz	400 Hz
	Mid × Hi	1000 to 10000 Hz	2500 Hz
Compressor: evens out a take that swings between quiet and loud	Thresh	−60 to 0 dB	−24 dB
	Ratio	1:1 to 20:1	4:1
	Attack	1 to 300 ms	6 ms
	Release	10 to 1000 ms	150 ms
Distortion: from light grit to heavy	Drive	0 to 1	0.3
	Mix	0 to 1	0.5
Delay: repeats, fed back on themselves	Time	10 to 1000 ms	250 ms
	Feedback	0 to 0.95	0.3
	Mix	0 to 1	0.25
Reverb: puts the line in a room	Decay	0.1 to 10 s	1.8 s
	Pre	0 to 300 ms	20 ms
	Mix	0 to 1	0.25
Limiter: a ceiling the level does not go over	Ceiling	−24 to 0 dB	−1 dB

Defaults are set so a newly added device is audible without wrecking the take.

Notes on specific devices:

• Limiter: put it last in the chain. Anything after it can push the level past the ceiling.
• Autotune runs a pitch detector on the incoming signal sixty times a second and feeds the result to a pitch shifter. Its default scale is Chromatic, which contains every note, so it pulls a flat note onto the nearest semitone without imposing a key. Use it when you do not know the key. Retune sets how fast the correction happens. Long values follow the singer's own drift and sound natural; at the bottom of the range the note snaps into place on the attack, which gives the well-known hard-tuned effect.
• The pitch detector uses normalized autocorrelation and ignores weak or noisy frames. Unvoiced consonants such as "s" and "t" have no pitch, so the shifter is left alone when the detector has no answer. This stops it jumping on sibilants.
• Reverb: changing decay or pre-delay re-renders the impulse response. The audio updates when you stop dragging; the reading on the card updates while you drag.

Parameter changes are ramped onto the running audio graph, so you can turn a knob while listening. Adding, removing or bypassing a device rebuilds only that lane's chain and does not re-schedule any clips.

Keyboard shortcuts

These shortcuts are bound to the window, so they work wherever focus is, with two exceptions:

• While you are typing in a text field, all of these shortcuts are off, so ⌘/Ctrl + Z in the script undoes your typing.
• Escape is handled by the lane area, not the window. Click in the timeline first.

Panels

Keys	Action
⌘/Ctrl + K	Command palette
⌘/Ctrl + R	Toggle the Script panel
⌘/Ctrl + P	Toggle the Prompt panel
⌘/Ctrl + B	Toggle the Library panel
⌘/Ctrl + E	Toggle the Sample panel
⌘/Ctrl + F	Toggle the Devices panel
⌘/Ctrl + I	Toggle the History panel
⌘/Ctrl + N	Toggle the MIDI panel
⌘/Ctrl + O	Toggle the Piano roll
⌘/Ctrl + G	Toggle the Logs panel
⌘/Ctrl + L	Toggle the Timeline

Holding Alt disables the panel shortcuts, so Alt combinations go to the browser.

Timeline

Keys	Action
Space	Play or stop
⌘/Ctrl + Z	Undo
⌘/Ctrl + Y, ⌘/Ctrl + ⇧ Z	Redo
⌘/Ctrl + A	Select every clip
Escape	Clear the selection; cancel an armed scissors
⌘/Ctrl + C	Copy the primary clip
⌘/Ctrl + X	Copy the primary clip, delete the whole selection
⌘/Ctrl + V	Paste at the playhead, on the copied clip's track
⌘/Ctrl + D	Duplicate the whole selection
Delete, Backspace	Delete the whole selection
M	Mute or unmute the primary clip
S	Split the primary clip at the playhead, when it is inside it
Alt (held while dragging)	Defeat snapping
Shift + drag the ruler	Mark a loop region
Shift + click a clip	Add or remove it from the selection
Wheel	Zoom about the pointer
Shift + wheel	Scroll horizontally

While an agent holds the story, undo, redo, playing and selecting still work. Every other editing action is paused.

When an agent rewrites the story, the undo stack is cleared, because the recorded undo steps may refer to clips that have moved or changed.

Generating into the timeline

The Prompt panel (⌘/Ctrl + P) is the same request form the Music studio uses, with the submit button labelled Generate & insert. Describe a sound or a piece of music, pick a model, and the result lands on this timeline instead of in a shelf.

Where it lands:

• In time: at the playhead if the story is playing, otherwise at the start of the selected clip, otherwise at zero.
• On which lane: the lane the device rack is pointed at, which follows your selection.

If the request asked for the take to be separated, all the stems come back as one take and each stem gets its own lane, stacked above the highest lane in use. Variations are alternatives, so only the first is inserted; the rest stay in the shelf for you to drag in.

Every attempt (running, done or failed) is recorded in the History panel (⌘/Ctrl + I). History is per story and survives a reload. It is a separate panel so the Prompt panel does not get shorter each time you generate.

See Generating for prompt technique and what each model wants.

Driving the studio with words

An instruction bar runs along the bottom of the whole studio. Type what you want done and press Enter (Shift + Enter for a newline). Use it for edits that would take many clicks, such as "name all of the tracks".

It shows you the plan before running it. Each edit is listed in plain terms: clips by their words or their sound's name, tracks by number and name. Destructive edits are marked. Apply runs the plan, Discard throws it away.

The bar uses the local prompt assistant, which is off by default. With nothing configured, the bar says so.

What works

The request is classified first, then planned by a separate step for its type. Roughly:

• Editing what is already there: "mute 1 and 3", "rename the tracks", "move the vocal to the cursor", "delete the last clip", "put a delay on the drums", "drop track 2 by 6 dB".
• Making something new for this timeline: "add some drums", "put rain under this", "add a bassline on track 2". This goes through the Prompt panel's own submit path, so the take arrives the same way a typed prompt's would and appears in the History beside it.
• Describing a whole finished piece with no reference to this timeline: "a hardcore instrumental where distorted guitars thrash over a double-kick kit". This is sent to the Prompt panel.
• Working the interface: "show the library", "open the script", "start recording".

References are matched by meaning: "the vocals" finds the track holding a clip labelled breathy female vocal hook. Track numbers are literal: "mute 1 and 3" means tracks 1 and 3, never positions in a list. "Here" and "at the cursor" mean the playhead; "these" and "the selected ones" mean your selection. Nothing has to be selected.

The action vocabulary

The assistant does not run anything itself and can only refer to clips and tracks it was sent. It returns a list of actions. The page applies each one through the same handler a click uses, so assistant edits update the page immediately and go on the undo stack like manual ones.

Action	What it does
rename_track	Give a track a name
set_track	Mute, solo, or set a track's gain
move_clip	Move a clip in time, onto another track, or both
set_clip	Mute a clip, or rewrite its words
delete_clip	Remove a clip. The one action flagged destructive
add_clip	Write a new line onto a track at a time
add_device / remove_device	Put a device on a track's rack, or take one off
generate_clip	Generate audio from a prompt and place it
show_panel	Open or close a panel
start_recording	Start a take

Every action that names a clip id or a track index is checked against the timeline that was sent. Actions that refer to something that does not exist are dropped before you see the plan, and the bar tells you how many were dropped. Device kinds are checked against the eight the rack can build.

This part of the studio is under active development. Rely on the actions above, not on any particular phrasing.

Export

Export mixes the arrangement down to one file in your library:

• Clips are placed at their real start times; gaps stay silent.
• Clips on different lanes sum, with crossfades applied where clips overlap on a lane.
• Clip gain and track gain both apply; muted clips and un-soloed tracks are left out.
• Trims are applied, so a shortened clip is short in the file too.
• Devices are not applied. See the rack, above.

The result is a normal sound. It appears in your library with the story's title and can be dropped onto another timeline.

For agents (MCP)

An agent connected over MCP can build and edit a story without the page being open. While an agent is working, the studio shows a notice and pauses editing; the work usually takes seconds. The agent's changes stream into the open page as they happen, so you do not need to refresh.

The studio-relevant tools:

Tool	What it does
soundbox.create_story	Start a story. Sets the default delivery for every line
soundbox.list_stories	Stories, most recently worked on first
soundbox.get_story	Every line: text, voice, timing, and whether it has been voiced
soundbox.add_lines	Append lines: spoken lines, pauses, or existing library sounds
soundbox.update_line	Words, voice, delivery, start, track, gain, mute, and trim
soundbox.delete_line	Remove one line. Its audio stays in the library
soundbox.split_line	Cut a voiced line at a point in its own audio
soundbox.speak_story	Read every line that has no audio yet. This is the slow one
soundbox.open_in_studio	Put a library sound on a timeline, split into stems where possible
soundbox.mix_track	Name, mute, solo or set the level of one lane
soundbox.remove_track	Delete a lane and the clips on it. Lanes above slide down
soundbox.export_story	Mix the arrangement down to one sound

Changing a line's words or delivery does not re-read it. Voice the line again to hear the change.

Splitting a voiced line keeps all the words on the first half, because there is no reliable way to tell which words fall on each side of the cut.

An agent can change the same track properties the export applies: name, mute, solo and gain. There is no MCP tool for the device rack, because the export ignores it.

The studio runs on an ordinary HTTP API: /api/stories, its /tracks and /segments, and /export. See the API reference to call it directly.