JingmeiAn AI group pet for your chats

Discord and Telegram · several characters from one config · built on Pi

Reads the room

Always answers when mentioned, replied to or called by name; otherwise chimes in by probability and content, gated by cooldowns and share of the conversation.

🖼️

Sees images and video

Up to 4 images per message go to the model, videos are sampled with ffmpeg, and a separate vision model can describe images for text-only models.

🎙️

Voice, drawing, long-text images

Fish Audio voice messages, Antigravity drawing, long replies rendered as images with formulas and plots, and live candlestick charts.

🧠

Remembers members

Names, birthdays and lasting facts people state about themselves, recalled on demand; each character keeps its own soul. Anyone can /forget.

🔎

Web and history search

DeepSeek web search with sources; older chat is found again by keyword and semantic search, scoped to the current group.

💸

Token-frugal

A stable system prompt and tool list hit provider prefix caches, conversation segments are bounded, and reasoning is off by default.

🛡️

Sandboxed math

run_js computes in a short-lived bubblewrap-isolated subprocess; without the sandbox the bot refuses to start.

🧰

Easy to keep

An init wizard writes the config, doctor checks everything, config edits hot-reload, and update is one command.

Up and running in a few commands

You need Linux, Bun 1.3+, bubblewrap, a Discord or Telegram bot and model credentials. See Quick start.

bash

git clone https://github.com/beiwater/jingmei.git
cd jingmei
bun install
bun run jingmei init   # wizard: platforms, tokens, characters, model
bun run start