MPC Live II · MPC OS 3.9.1.2 · armv7
The device
Fourteen facts, each of which cost real time. They are not opinions; they were read off the device or learned the hard way.
1 · It is a Linux box with a BusyBox userland
ssh root@<ip> # root, LAN only
ps | head -n 5 # BusyBox: "head -5" does not exist
systemctl status acvs # the MPC app runs as acvs.service
acvs is the whole application. Restart it and the screen goes down and comes back;
that is how a new .so is loaded. Older firmware calls the unit inmusic-mpc.
2 · A plugin is a folder and one XML line
/sdcard/Synths/<vendor> - VST - <name>/
mpc_<something>.so
version.xml
Plugin Skins/{TUI.json, Q-Links.json, *.png}
MPC.settings → pluginList-arm → KNOWNPLUGINS → <PLUGIN format="VST" file="…"/>
MPC finds a skin by folder name: <vendor> - VST - <product>. The
.so may live elsewhere; the folder name is what binds.
3 · Never edit MPC.settings while MPC runs
It keeps settings in memory and rewrites the file. A malformed edit resets everything to defaults.
The only correct sequence is stop → back up → edit → start. Every installer and
mpc-store.sh does exactly that; there is no shortcut worth the risk.
4 · A new skin does not need a restart; a new .so does
MPC re-reads Plugin Skins/TUI.json on every insert, so remove the instance and insert
again to see a changed skin. The shared library, however, stays mapped for the life
of the process: replacing the file on disk does not change the running code, and a remove + re-insert
does not reload it either. Check with grep multihead /proc/$(pidof MPC)/maps — the entry
reads (deleted) when the file was replaced under a running process, and the plugin you
are looking at is still the old build.
5 · A thread that outlives the plugin can take the app down
Two “OOPS! Your MPC recovered from an internal error” in one afternoon came from a
listener thread the plugin started and never stopped. Whether the host actually unloads the library
is still not proven on this firmware — we measured that it does not — so the honest rule is
the defensive one: a worker thread must be stopped and joined before
destroy_instance returns. It costs nothing and closes the case if the host ever
does unload. A dlclose test reproduces the failure shape offline.
6 · The screen is a PNG canvas you cannot draw on
The plugin area is 1280×628 inside a 1280×800 screen. Layout coordinates are shadow coordinates:
a widget’s cy is drawn 86 px higher. Below y ≈ 542 there is no room for widget labels any
more — bake those into the plate. MPC draws every label and value; you supply the picture and the
values it polls.
7 · Widget animation costs a redraw
A busy screen has audibly crackled the audio on this hardware. A single meter strip polled every ~46 ms is fine. Anything more belongs off the device — that is the whole reason for the companion screen.
8 · The host has no “reload” and no “unload”
There is no API to reload a plugin, so a build–see-it loop is: deploy, remove the instance, insert
it again — and restart acvs when the .so changed. Everything else is
waiting for a human to tap a plugin list.
9 · Screenshots and input come from an addin, not from the app
jacob-sabella/mpc-addin-remote (MIT, LD_PRELOAD) serves the live screen,
touch injection, MIDI and an MCP endpoint on 127.0.0.1:6720:
screenshot, tap, drag, scroll,
wait_for_screen, play_notes, device_status,
list_files. Hardware buttons are not injectable (no virtual RawMIDI port), so
anything behind PROJECT → LOAD is still a human’s job.
Argument shapes matter and fail silently: drag takes flat
from_x/from_y/to_x/to_y, and scroll needs
x, y, direction and an amount of 20–2000 px.
A wrong shape looks exactly like a tap that did nothing.
10 · Files that carry state
| Path | What |
|---|---|
/sdcard/Synths/… | plugin folder (skin, .so, presets) |
/sdcard/MPC Documents/Plugin Presets/Insert Effects/<folder>/ | effect presets (.xpl) — not the skin’s Presets/, which is the instrument path |
/sdcard/MPC Documents/Projects/*.xpj | songs; a half-written one is a lost session |
/data/Settings/MPC/MPC.settings | the plugin registry (see 3) |
/data/sentry/.sentry-db/ | crash records; the payload has no symbols |
11 · A preset menu that stays empty
A port without "presets": "presets.json" in vst.json has an empty PRESET
field even when .xpl files sit in the right folder: the field is the VST program list, not
the file list. Each preset needs at least one value or the build aborts.
12 · Parameters are read back as text
effGetParamDisplay is what MPC shows, and it is also what the host compares when it
decides whether a control moved. Get the text wrong and the control springs back — the sound changes,
the display does not, and it looks like a broken button.
13 · CPU is measured, not estimated
tools/bench.sh plays the plugin on the device and times the audio thread:
idle, audio in, Q-Link sweep, release tail, one-shot triggers separately. Verdict against the
2902 µs block. Numbers from a desktop are meaningless — an M-series core is not a Cortex-A17.
14 · What you cannot get for free
- Audio capture off the device: there is no loopback. A plugin can report its own levels; anything else is a microphone.
- Crash symbols: the built-in crash records carry no frames, so a crash is
diagnosed by correlation and by
/proc, not by a backtrace. - Hardware buttons: see 9.
- A second screen: see 6 and 7.