liguozheng-ac/Lingbo-Desk

What it does

Lingbo Desk is a Windows desktop agent built around continuous voice conversation. The desktop client shows a Live2D avatar whose mouth and basic motions follow speech, and it lets the user interrupt the current reply or stop old playback. Speech recognition and synthesis support Chinese, Japanese, and English, while chat runs on DeepSeek's cloud API or a local Ollama model (6).

Beyond conversation, the agent can work in the background while the user keeps talking: it handles files and Office documents, drives an independent browser session, reports task progress, and supports pausing, modifying, cancelling, and resuming checks. It also stores explicitly saved long-term memories and imports versioned Skills (6).

The architecture is an Electron/Vue 3/TypeScript desktop client in front of a local Node.js backend that exposes HTTP APIs for models, speech, tasks, permissions, and user configuration. An Agent worker process runs the tool-call loop; tool requests return to the main service for validation and execution (13). File and browser operations are constrained by directory authorization, action confirmation, file hashes, backups, and execution receipts.

Getting started

Requirements are Node.js 24, pnpm 11, and Python 3.12. The repository does not contain secrets, chat history, or installed dependencies, so first-time configuration, dependency installation, and avatar asset import are covered in the setup documentation (15).

A source checkout can be prepared as follows from the repository root (18):

pnpm --dir desktop install --frozen-lockfile
pnpm --dir engine install --frozen-lockfile
pnpm --dir speech install --frozen-lockfile
python -m venv .venv
.venv/Scripts/python -m pip install -r requirements.txt
Copy-Item .env.example .env
# edit .env with your own service credentials
pnpm start

The local backend listens on http://127.0.0.1:3850 by default. Building the desktop app additionally requires importing a Cubism Core SDK copy whose terms you have accepted; the source repository ships neither the SDK binary nor a character model pack (19):

pnpm --dir desktop run build:modules
$env:LIVE2D_CORE_PATH="C:/your-sdk/Core/live2dcubismcore.min.js"
pnpm run prepare:desktop
pnpm run build:desktop

A Windows x64 portable test package is described separately in the Windows test-package notes. Maintenance checks include pnpm test, pnpm run verify:task, pnpm run check:source, and a desktop typecheck (22).

Constraints

Long-term memory requires explicit saving; automatic extraction and vector retrieval are not implemented. Two task pipelines are still in use for general work and staged reports. A local Ollama model does not make cloud speech automatically offline. The Agent worker is not an operating-system sandbox, and imported Skills scripts run with the current user's permissions. Complex Office documents, long-running sessions, sleep/wake behavior, and cross-device deployment still need further verification (24).

Lingbo's own code is MIT-licensed, but bundled third-party code follows its own licenses, and the original Live2D preview image requires separate permission for reuse (25).

HotRepo's take

The repository has no recorded GitHub daily ranking. In HotRepo's new-project snapshot for TypeScript on 2026-11-07, it showed 182 stars and 53 stars gained over a window of roughly 23.5 hours; the star curve moved from 102 to 182 between 2026-11-06 and 2026-11-07. That is a sign of interest in a newly created project, not evidence of maturity. The README is unusually candid about incomplete areas — the two task pipelines, the lack of an OS sandbox for Skills, and pending Office and sleep/wake verification — and those points matter more for adoption than the star count.