lujingshi1978-glitch/ds-life-and-soul

What it is

Feiyu Life is a local, persistent AI agent that runs on a locally deployed DSH (DeepSeek Harness) instance. The project ships the agent runtime, native DSH plugins, a "life console" dashboard and desktop entry points in one deployment. Its design centers on keeping experiences, interests, relationships and current thoughts in separate local stores and recalling them by situation, rather than treating every conversation as stateless chat.

It also connects web pages, video, files, WeChat text input, QQ and desktop chat through the same character memory and action boundaries. The console lets you inspect inner state, activities, memory, relationships, knowledge, character settings, connections, permissions and usage feature documentation.

Quick start

For end users, the README points to the Windows one-click 0.6.0 package (about 1.04 GiB). Download and fully extract it, then double-click 肥鱼.exe, fill in the API settings, click "安装并打开", and log in to any platform accounts you need. Node, Python and DSH are bundled; you do not need to install them separately.

For source mode, the currently validated platform is Windows 10/11 x64:

1. Install Node.js 24 LTS (with npm). Shortcut voice also needs Python 3.11 plus a microphone and audio output.
2. Extract to a short path such as D:\Feiyu.
3. From the repository root, either double-click 启动肥鱼.cmd or run:

node scripts/deploy.mjs launch

On first run the installer asks for options, installs dependencies, creates an isolated feiyu DSH profile, generates a new character profile and local credentials, and then starts the life console. Configure the model provider and key in the local DSH interface; platform accounts must be logged in by you. Default ports are DSH 3081, life console 3110 and desktop gateway 3117; the launcher refuses to take over ports that belong to other instances installation details.

The shortcut voice hotkey is Ctrl + Alt + Space: press to start recording, press again to stop (silence can also end it). Recognized text and replies appear as small subtitles instead of a large chat window; the tray lets you adjust transparency, size, position and mouse pass-through.

Defaults and boundaries

A fresh install imports no historical memory. OneBot/QQ is off by default with an empty whitelist; proactive contact and independent website interaction are also off, as are restricted creation and independent work entry points. The shortcut voice grants no extra permissions to websites or QQ. The daily cost soft limit defaults to 0, meaning disabled; suggested presets are 3, 10 and 20 yuan for saving, daily and rich memory settings, but these are scheduling preferences based on your configured prices, not provider bills or hard quotas default behavior.

Reaching the limit pauses new autonomous activity and high-cost work while bounded work finishes and normal incoming messages can still be answered; in-flight requests are not killed. If costs cannot be verified when a limit is enabled, new background activity is paused and the page says so.

Data and privacy

Local speech recognition uses a fixed Whisper model: source mode downloads it on first explicit voice installation, while the Windows one-click build carries it. Recognition reads only local files. Default speech output uses Edge TTS, so the text to be spoken is sent to that service; conversation text goes through DSH to whichever model provider you choose README. Platform accounts, including QQ gateways, are configured separately, and the package does not distribute or auto-start gateways such as SnowLuma.

HotRepo's take

The project appeared on HotRepo's new JavaScript list on 2026-11-02, recording 3 stars in the roughly one-hour growth window from 22:45 to 23:45 UTC, moving from 103 to 106 total stars. That is a single short snapshot, not a sustained trend, and the repository has no recorded GitHub daily rank. The notable engineering choice is packaging a 1 GiB Windows installer that bundles Node, Python, DSH, a voice model and fallback/upgrade logic, which makes first-run setup easy but also concentrates trust in a prebuilt binary. The README states the one-click package went through acceptance on clean Windows 11 x64 and links a release record and correspondence scope, which is the right place to verify before running it.