lh-vocaloid6-mcp

Get a program inside a running VOCALOID 6, holding its own object graph, so it can read and (once calibrated) change the open song — with no clicking and no typing.

VOCALOID 6 has no scripting API, and unlike CeVIO its engine is not a managed library that can simply be referenced: the note model is native code inside the editor process. So the bridge is a DLL loaded into VOCALOID 6 which hosts an MCP server in-process and publishes it on loopback HTTP. Not screen scraping, not keyboard, not mouse.

One C++ source file builds two artifacts:

The bridge is the same source as the VoiSona one, built with the VOCALOID 6 identity; the tool prefix and the engine guard differ, nothing else.

Read this before you plan anything

Two measured facts change what is worth attempting here. Both come from injecting the bridge into a real VOCALOID 6 (6.13.0.1) and looking.

1. VOCALOID 6's editor is WPF on .NET, not a native C++ editor. Its host image is only 0x3d000 bytes, and its windows are HwndWrapper[VOCALOID6;;...] — the editor is a shell over managed assemblies.

2. A note map is a chain of raw pointers read from a module base, and the GC moves managed objects. A chain that resolves today can go stale under you, and a stale chain does not fail: it reads plausible garbage and writes it back.

Therefore: do not start a note-map discovery pass on VOCALOID 6 until someone has settled whether a stable chain exists at all. VoiSona is JUCE and native and fits that workflow as written; VOCALOID 6 does not, yet. capabilities.notes will read "none" and that is the honest state, not a missing step.

Related MCP server: ce-mcp

Requirements

• Windows, VOCALOID 6 installed, and the editor running before you attach.

• Python 3.10+.

• ziglang — the only toolchain the native bridge needs.

Install

git clone https://github.com/ppzbr/lh-vocaloid6-mcp
cd lh-vocaloid6-mcp
python -m pip install -e ".[mcp]"
python -m pip install comtypes
python -m pip install ziglang

Build the bridge:

powershell -ExecutionPolicy Bypass -File bridges/native/build.ps1

It writes lhai_bridge.build-info.txt beside the artifacts. That record, not the compiler, is the proof of what was built. If the build fails, stop and report the compiler output.

Prove the transport with no editor involved

Run lhai_bridge.exe and drive it over HTTP:

initialize   -> serverInfo LhBridge, protocol 2024-11-05
tools/list   -> the vocaloid_* surface

This step exists so that a later failure is known to be about the host, not about the bridge.

Attach, and know that it worked

bridge_inject(engine="vocaloid6")

What a healthy injection looked like on the development machine (the shape matters, not the counts):

host image in the module list     255 modules
scan "4D 5A"                      found at the image base
chain u16 at the image base       0x5A4D  ("MZ")
read at the image base            4d5a90000300000004000000ffff0000
windows                           the editor's own HwndWrapper[...] windows
status                            level: partial, capabilities.notes: "none"

Then the surface:

bridge_list()                      # every bridge answering right now, and its engine
bridge_status(engine="vocaloid6")  # version, app baseline, per-capability levels
bridge_tools(engine="vocaloid6")   # the tools with their real argument schemas
bridge_call(engine="vocaloid6", tool=..., arguments=...)

The engine-prefixed tools are the memory primitives — modules, windows, scan, read, write, chain, load_map, status. chain resolves exactly as a map file states it: p = base; for each offset o: p = *(void**)p + o, then read the typed value at p.

The wrong-engine guard is deliberate: a VOCALOID 6 bridge rejects a voisona_* tool name, and vice versa. That answer means you are talking to the wrong bridge.

Detach

bridge_eject(engine="vocaloid6")

Always eject before rebuilding. Windows will not overwrite a DLL loaded in a running process.

Rules this bridge lives by

1. Do not modify, replace or delete any VOCALOID 6 file.

2. Never kill the editor process.

3. Never substitute UI automation, keystrokes or clipboard for a bridge capability.

4. If a step fails, stop at that step and report exactly what happened.

What is verified, and what is not

• Verified: the bridge loads into a real VOCALOID 6 (6.13.0.1), publishes its endpoint, and answers the process, module, scan, read, chain and window primitives — the MZ evidence above is from that run.

• Not verified, and deliberately not attempted: any note-level read or write. The WPF/GC facts above are why. The bridge will tell you capabilities.notes: "none".

• The fallbacks that do work today: the project-file route, and the UIA live layer.

When something looks wrong

Layout

The client ships with adapters for the other editors in the family (lhai/mcp/engines/). They are inert unless that editor is the target.

Tests

python -m pip install -e .
python -m pytest

Licence

MIT. See LICENSE.