Three letters, and what is under them
MIDI Polyphonic Expression is ordinary MIDI with the addressing changed: every note gets a channel to itself, so bending, pressing or brightening one note leaves the others alone. Sonari sends it. This page says what leaves, message by message.
Ordinary MIDI says that a note started and how hard, and then addresses everything afterwards to the channel rather than to the note. One pitch bend, one pressure, one brightness, shared by everything sounding there. Hold a chord, bend it, and the whole chord bends. No synth is doing that wrong: it is what the message says.
On a keyboard that is honest enough, because a piano cannot bend one key either. On a guitar, or a voice, or a hand in the air, it is half of every note missing. What a string does after it is struck is not what the string beside it is doing, and until 2018 there was no agreed way to say so.
It adds no new message. It changes the address. A block of channels is set aside and treated as one instrument: one channel carries anything meant for the whole block, and the rest carry one note each. A note takes the next free channel when it starts and gives it back when it stops, so a bend, a pressure or a brightness arriving on a channel is arriving at exactly one note.
Three dimensions per note, and every MPE instrument has the same three: how far the note is bent, how hard it is being pressed, and a third axis that most instruments spend on brightness. The third one travels as controller 74 and is usually called slide.
Because none of it is a new message, a synth that was written before any of this understands each one perfectly well. The MIDI Association adopted it into the specification in 2018.

Two hands raised in an empty concert hall, fingers half closed, turning and moving independently of one another.
Logic Pro. Alchemy and Sculpture are the two people usually reach for. The step that is easy to miss is setting the track’s MIDI channel to All: a track listening on one channel cannot hear notes arriving on the other fifteen, which is what silence looks like when everything else is right.
Ableton Live. The switch is on the port rather than on the track. Wavetable and Sampler answer per note.
Everywhere else it is usually the instrument plug-in that has to be put in MPE mode, and the host only has to pass the channels through. The setup for each is on the DAW page.
Sonari is a Mac application that watches you through the camera the Mac already has, and it publishes one MIDI source that anything else on the machine can listen to. With MPE on, this is what goes down it.
Before any note. Channel 1 is announced as the one that speaks for the group, with the standard configuration message, and the fifteen channels above it are handed to notes. Then a bend range is declared on each: two semitones on channel 1 and forty eight on the note channels, which is the figure MPE instruments conventionally use.
A note on, when the hand closes. Where the hand is across the picture chose which note. The loudness is how closed the hand already was at the moment it landed, because a camera cannot see a strike. The hand has to close about a third of the way to a fist for it to sound, and open past a lower mark before it lets go, so a hand resting on the line does not machine gun the note.
Pitch bend, on that note’s own channel, from the roll of the wrist. About four semitones either side of centre rather than the whole forty eight, because a quarter turn of the wrist is the whole axis: a wobble is vibrato and a deliberate roll is a bend. A wrist at rest sits at its neutral roll, so a hand doing nothing is in tune.
Controller 74, on that note’s own channel, from the height of the hand. Up is brighter, which is the direction every instrument that reads it expects.
Channel pressure, on that note’s own channel, from the squeeze. Measured from the point where the note started rather than from an open hand, so the whole travel of a squeeze is the whole range of the pressure.
A note off, when the hand opens. The channel goes back.
One more, and it is the part that is usually wrong elsewhere: a hand with nothing sounding sends no expression at all. It does not fall back to the channel that speaks for the group, because a message there is addressed to every note in it, and a silent hand’s wrist roll would then bend the other hand’s note.
None of this is confined to one style. Turn MPE on and notes from any of the 17 ways to play take a channel each. And Sonari’s source is a MIDI 2.0 source, so where a host takes MIDI 2.0 the same movements arrive at more than seven bits of resolution and the brightness can be addressed at the note itself; where a host does not, macOS converts on the way out and what arrives is the MPE above.

Sonari playing chords from a Mac keyboard: the chord changes, the keys light as each note sounds, and the readout names what is sounding.
Two voices, not ten. A keyboard you touch gives you as many notes as you have fingers. Here it is one note per hand, because about four separable continuous quantities per hand is what a camera can honestly measure, and reading fingertips instead would buy more voices at the price of one moving quantity each.
No strike. Nothing measures how fast a key went down, because there is no key. Loudness at the start of a note is how closed the hand already was.
Nothing to touch. There is no surface, so there is no edge to steady a finger against, which is the price this whole family of instruments has been charging since the theremin.
The ones people mean when they say MPE controller, at the time of writing and rounded. None of them is a camera and all of them are surfaces you touch.
Haken Continuum, around $1,799 to $6,510 depending on its length, with the ContinuuMini around $999. Expressive E Osmose, around €1,799 to €1,999. LinnStrument, around $1,149 to $1,599. Embodme Erae II, around €849. Intuitive Instruments Exquis, from around $275.
ROLI Seaboard RISE 2, around £1,099, and this is the one to be careful about: ROLI owned this subject, and at the time of writing roli.com returns an error page. There is a page here for anybody who was waiting on one.
What the list says is that an MPE instrument is a made object from a small company, and priced like one. The cheap camera-based tools are a different trade and are compared on the gesture MIDI controller page.
It is not a controller you buy and plug in. It is a Mac application, it uses the camera that is already in the machine, and it makes its own sound: 10,000+ of them. Its own instrument reads the same three dimensions the wire carries, so the press, the brightness and the bend are audible before any of this reaches a host at all.
And then it will just send, to whatever you already run, on the same Mac, with nothing to plug in and no driver to install.
MPE exists so that a note can go on being played after it has started. Sonari sends it from a hand that never touched anything, which is either the whole point or a good argument for buying a keyboard.
Sonari is an instrument you play in the air, on the Mac you already have. Free for fourteen days, and nothing to set up first.