[](https://github.com/shaobeichen/dsh-im-bridge/stargazers)
[](https://github.com/shaobeichen/dsh-im-bridge/network/members)
[](https://github.com/shaobeichen/dsh-im-bridge/actions/workflows/ci.yml)
[](LICENSE)
**Command DeepSeek Harness from Feishu, WeCom, or Telegram: remote tasks · result notifications · risk approval**
[中文](README.md) | [English](README.en.md)
**dsh-im-bridge** is an IM channel for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness).
DeepSeek Harness (DSH) is an AI agent engine that runs locally on your computer — it can read and write code
in your working directory, run commands, and execute long tasks. Normally you control it through its web interface.
dsh-im-bridge adds a "remote control" to that interface: you message a bot in Feishu, WeCom, or Telegram,
those messages become instructions for DSH, and results come back to your chat.
In short: **your chat window is DSH's control panel**.
## ✨ Supported platforms
| Platform | Status | Notes |
|---|---|---|
| Feishu (Lark) | ✅ works | Official long connection, approval via card buttons, verified live |
| WeCom (WeChat Work) | ✅ works | Callback URL + message encryption, approval via text command |
| Telegram | ✅ implemented | No public webhook needed (polling), approval via buttons (bring your own bot token) |
| DingTalk | planned | — |
## 📦 What's inside
- **Chat control**: send tasks to the AI from IM, as naturally as chatting
- **Risk approval**: dangerous operations (file deletion, system commands) require your confirmation — denied by default
- **Result notifications**: tasks pushed on completion or failure, with duration and token usage
- **Full output export**: when long output is truncated, use `/log` to send the full text as a file
- **Session management**: `/new` to start, `/status` to check, `/mute` to silence notifications, and more
- **Access control**: configure who can use it and who is an admin (only admins can approve and change settings)
## 🚀 Getting started
**Prerequisite**: [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) installed on your computer. If `dsh` reports `command not found`, install it first:
```sh
npm install -g @deepseek-ai/dsh # global install; verify with: dsh --version
# Don't want a global install? Prefix every command with npx: npx @deepseek-ai/dsh
```
All four platforms follow the same path: **install the plugin → configure credentials → restart → chat**. No need to clone this repo or write code.
> Want all channels at once (or one-click install from dshmarket)? `dsh plugin --profile web add github:shaobeichen/dsh-im-bridge` is equivalent to installing the core + Feishu + WeCom + Telegram + WeChat together; channels without credentials just show "disconnected" and don't affect the others.
| Platform | One-command install | Prepare | Use |
|---|---|---|---|
| Feishu | `dsh plugin --profile web add dsh-im dsh-im-feishu` | **Web QR setup (recommended)**: Settings → Plugins → Feishu → "Scan to connect", then scan with Feishu (or terminal `npx -y dsh-im-feishu-qr`); manual app creation: [guide](docs/feishu-setup.md) | Configure credentials, restart `dsh web`, private-chat the bot: `/new` → send tasks; risky operations send an approval card — tap a button |
| WeCom | `dsh plugin --profile web add dsh-im dsh-im-wecom` | Create a self-built app in the [admin console](https://work.weixin.qq.com/wework_admin/frame), get CorpID / AgentId / Secret, configure callback URL and Trusted IP ([guide](docs/wecom-setup.md)) | Configure credentials, restart `dsh web`, open the app from "Workbench": `/new` → send tasks; approval comes as text — reply `/approve yes` (no buttons) |
| Telegram | `dsh plugin --profile web add dsh-im dsh-im-telegram` | Create a bot with [@BotFather](https://t.me/BotFather), get the token ([guide](docs/telegram-setup.md)) | Configure credentials, restart `dsh web`, private-chat the bot: `/new` → send tasks; risky operations send an approval card — tap a button |
| WeChat | `dsh plugin --profile web add dsh-im dsh-im-weixin` | **Web QR bind (recommended)**: Settings → Plugins → WeChat → "Scan to bind", then scan with WeChat (pairing digits entered in the page); prerequisite: your account has the WeChat bot entry under Me → Settings → Plugins ([guide](docs/weixin-setup.md)) | Restart `dsh web` after binding, then message the bot in WeChat: `/new` → send tasks; approvals come as text — reply `/approve yes` (no buttons) |
> Credentials go in environment variables (`FEISHU_APP_ID`, `WECOM_CORP_ID`, `TELEGRAM_BOT_TOKEN`, etc.). See each platform's setup doc for details.
> [!NOTE]
> Want to run the integration scripts **without installing into DSH** (requires cloning the repo + Node.js 22+)? See the demo runners under "For developers" below.
## 🗂 Package layout
| Directory | Description |
|---|---|
| packages/im | Core: message handling, sessions, approvals, notifications |
| packages/im-telegram | Telegram adapter |
| packages/im-feishu | Feishu adapter (incl. QR setup `dsh-im-feishu-qr`) |
| packages/im-wecom | WeCom adapter |
| packages/im-weixin | WeChat adapter (Tencent iLink, incl. QR bind `dsh-im-weixin-qr`) |
| demo | Demo and runner scripts |
| docs | Documentation |
## 🧭 Architecture
Message flow:
```
Message from your IM
↓
Channel adapter (converts platform message to a unified format)
↓
Core dsh-im (command or task? session management, approvals)
↓
DeepSeek Harness (does the actual work)
↓
Result sent back to your chat
```
For more, see the [docs index](docs/README.md).
## 🤝 For developers
- Want to modify code or add a channel? Start with [CONTRIBUTING.md](CONTRIBUTING.md)
- Repo rules for AI agents: [AGENTS.md](AGENTS.md)
- Original product spec: [docs/PRD-v0.5.md](docs/PRD-v0.5.md)
**Run the integration scripts without installing into DSH?** (clone this repo + Node.js 22+)
```sh
node demo/mock-demo.mjs # terminal-only mock IM
node demo/feishu-real.mjs --mode demo # live Feishu (needs FEISHU_APP_ID/SECRET)
node demo/wecom-real.mjs --mode demo # live WeCom (needs WECOM_* credentials)
node demo/telegram-real.mjs --mode demo # live Telegram (needs TELEGRAM_BOT_TOKEN)
```
## 🔒 Security
This project executes remote instructions and touches real files, so security is a priority. The core principle: denied by default, risky operations require your approval. Security model and vulnerability reporting: [SECURITY.md](SECURITY.md).
**End users don't configure anything security-related**: everyone is denied by default, and a stranger's first message triggers a trust confirmation — the admin (a user listed in `security.admins`) taps ✅ or replies `/trust` to let them in, permanently. The admin configures their own key once.
## 📄 License
MIT
