Hesi

<p align="center"> <strong>Hesi: Universal Terminal Center in Browser</strong><br> <em>Where multiple AIs meet to think together</em><br> <em>Run any CLI, connect any Agent, control any browser</em> </p> <p align="center"> <a href="https://github.com/qiuqiukof-oss/Hesi/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue.svg" alt="License"></a> <a href="https://nodejs.org/en/"><img src="https://img.shields.io/badge/node-%3E%3D18.0.0-brightgreen" alt="Node Version"></a> <img src="https://img.shields.io/badge/tested_on-Windows%20%7C%20macOS%20%7C%20Linux-lightgrey" alt="Platform Support"> </p> <p align="center"> <img width="2549" height="1191" alt="image" src="https://github.com/user-attachments/assets/0ee0f21a-d3f9-49f8-a8a6-0e605aafe11a" /> </p> <p align="center"> <b>Turn your browser into your development command center</b> — Multi-session terminal, multi-Agent round-table collaboration, browser automation, and MCP services, all self-hosted and ready to use. </p> --- > 🌐 **English documentation available:** [README_en.md](./README_en.md) > ⚠️ **Security Warning** > > Hesi can execute arbitrary terminal commands and control browsers through WebSocket (CDP integration). > **It is strongly recommended to use it locally on loopback addresses (`127.0.0.1` and `::1`) only. If you must expose it to a non-local network, be sure to set `QCLI_ACCESS_TOKEN` and read the "Secure Deployment" section below.** > Public deployment without authentication may lead to remote code execution (RCE) risks. --- ## 🚀 30-Second Quick Start **Want to get started quickly?** Use the offline portable package, double-click to use, no need to install Node: | What do you want to do? | How to do it | |------|--------| | 💻 **Try without installation** (recommended for beginners) | Download the [Desktop Tray Bundle](#desktop-tray-bundle), double-click `tray.bat`(Windows)/`./tray.sh`(macOS/Linux), and your browser will automatically open `http://127.0.0.1:4264` | | 🛠️ **Run from source code** (for developers) | `git clone` → `npm install` → `npm run build` → `npm start`, open `http://127.0.0.1:4264` | | 🤖 **Let AI work for you** | In "Settings", fill in `OPENAI_API_KEY` or `ANTHROPIC_API_KEY`, and then directly chat in the chat box ( **you can still use the terminal without filling in** ) | > 💡 You will see a new user guide bubble on the first opening. If you don't know what to do, try: **Open a terminal tab → Enter `dir`(Win)/`ls`(Mac/Linux) → Ask AI in the chat box "Explain the current directory"**. 📖 **Complete Beginner's Guide**: [docs/getting-started.md](./docs/getting-started.md) ｜ 🧑‍💻 **Contributor's Guide**: [CONTRIBUTING.md](./CONTRIBUTING.md) --- ## ✨ Why Choose Hesi - 🧭 **One tab, control everything** — Terminal, AI Agent, browser, document conversion, MCP tools, all unified in one web console. - 🤝 **Multi-Agent round-table collaboration** — Let opencode, Codex, aider, and other CLI Agents have a multi-round "round-table discussion" on the same task, colliding solutions instead of fighting alone. - 🛡️ **Headless execution, no conflict** — CLI Agents run headlessly in the background, rendering and your interactive terminal completely decoupled, with a clean and non-interfering experience. - 🔌 **Open and scalable** — Modular MCP server + plugin system + CLI preset templates, connect to your toolchain as needed. - 💻 **Desktop tray portable version** — Comes with an offline Node runtime, double-click `tray.bat` to use immediately, zero installation. - 🔒 **Secure and auditable** — Session-level authentication, audit logs, and security policies, with default zero configuration on local loopback. --- ## Table of Contents - [✨ Why Choose Hesi](#✨-why-choose-hesi) - [Overview](#overview) - [The Name](#the-name) - [Features](#features) - [🚀 30-Second Quick Start](#-30-second-quick-start) - [Quick Start](#quick-start) - [Desktop Tray Bundle](#desktop-tray-bundle) - [Architecture](#architecture) - [Tech Stack](#tech-stack) - [Testing](#testing) - [Contributing](#contributing) - [Secure Deployment](#secure-deployment) - [License](#license) --- ## Overview **Hesi** is a web-based universal terminal bridge platform. It combines `node-pty` + `xterm.js` with WebSocket real-time communication, allowing you to experience a native terminal in your browser. On top of this, it integrates AI dialogue, Agent management, MCP modular services, visual panels, browser automation, and more. ### Core Philosophy | Concept | Description | |------|------| | **Everything is CLI** | Launch any command-line tool in your browser, with multiple tabs running independently | | **AI Native** | AI assistant integrates terminal context awareness + tool call chain (loop detection to prevent loss of control) | | **AI × CLI Agent Collaboration** | Let AI assistants and CLI Agents (like opencode, Codex, aider) have multi-round "round-table discussions" on the same issue, colliding solutions | | **Modular MCP** | Model Context Protocol service, supporting session persistence, security policies, and audit logs | | **Browser Control** | Control Chrome/Edge through CDP integration, supporting script injection, network monitoring | | **Scalable** | Plugin system + preset templates + theme customizer + custom CSS injection | --- ## The Name **HeSi** — A name, two layers of thought. **Chinese** - **He** — Collaboration and combination: multiple Agents gather in one place, with Western and Chinese tool ecosystems connected here. - **Si** — Thinking and inspiration: AI's intellectual activity, as well as that fleeting idea. - Together, "HeSi" is homophonic with "coincidental inspiration", with a bit of unexpected surprise. The two characters are memorable at a glance. **English** - **HeSi** — four letters, one keystroke away in your terminal. A name built for the CLI. - **He** + **Si** (pinyin "Si") happens to be read as *"He thinks"* — A thinking partner, purely a wonderful coincidence. - The ending vowel makes it smooth to pronounce. > In one sentence: **HeSi is where multiple minds meet to think — together.** --- ## Features ### 🖥️ Multi-Session Terminal - **xterm.js + WebGL rendering** — Smooth terminal output, supporting WebGL and Canvas dual rendering backends - **Multi-tab management** — Each tab has an independent PTY process, not interfering with each other, supporting drag-and-drop sorting and pinning - **Terminal search** — Ctrl+Shift+F search in terminal output, supporting up and down navigation - **Link recognition** — Automatically recognize file paths, URLs, and click to preview in the browser - **Adaptive width and height** — Adapt to container size through FitAddon - **Session persistence** — IndexedDB stores terminal content, restoring after page closure - **Font adjustment** — Ctrl+= / Ctrl+- real-time adjustment, supporting 8-32px range ### 🔍 CLI Auto-Discovery & Presets - **PATH scanning** — Automatically scan system executable files, generating a quick launch list - **Preset system** — Built-in presets for developers, data scientists, system administrators, media engineers, and more - **Preset inheritance** — Presets support `extends` chain inheritance, sharing basic configurations - **Cache acceleration** — 24-hour disk cache, avoiding repeated scanning on each launch - **Version detection** — Automatically detect CLI version numbers (trying multiple flags: `--version`, `-v`, `-V`) - **Type classification** — Automatically determine interactive/batch type, supporting custom classification - **Folder organization** — Support creating folder groups for management, drag-and-drop classification - **Favorites** — Left sidebar can collect commonly used CLI; discussion mode Agent dropdown and favorites **automatically synchronize** (favorite items default to checked, pinned, and marked with ★) ### 🤖 AI Integration - **Multi-provider** — Support OpenAI, Anthropic, LM Studio (local model) - **SSE streaming output** — Structured event types (token/status/error/tool_calls/usage), 120s idle timeout protection (configurable `HESI_LLM_STREAM_IDLE_MS`) - **Terminal context awareness** — Automatically capture the latest 100 lines of terminal output as a system message - **Incremental context pruning** — Only send changed incremental lines, saving tokens - **Tool call chain** — Continuous tool calls with loop detection (ring detection + window deduplication + hard limit), preventing "instant overflow" style loss of control; depth thinking panel presents real-time tool cards (running → completion time + result preview, XSS security) - **Lightweight self-inspection** — Identify "comprehensive self-inspection / self-inspection" intent, automatically set the tool round limit to 6 rounds, significantly reducing redundant LLM calls - **Tool set** — File read/write, web search, terminal execution, document conversion, image/video generation - **Browser control tools** — Navigation, screenshot, click, input, execute JS, DOM snapshot, form filling - **Self-evolution** — Can read/modify its own source code, rebuild frontend, screenshot check UI ### 🧠 Cross-Session Memory > Let AI remember you: conversations are no longer lost after refresh/restart, and can be **recalled across sessions**. - **Server-side persistence** — Each conversation is stored in `data/memory/` ("session" as the unit, not relying on browser localStorage, still recoverable after refresh/restart). - **Left column session list** — Chat panel left column adds session column: new / continue / search / double-click rename / delete; grouped by "today / yesterday / past 7 days / earlier". - **🧠 Memory drawer** — Top 🧠 button opens, displaying **automatically generated user profile** and **remembered facts**, each fact can be "forgotten". - **Automatic summary compression** — Long conversations exceeding the threshold are automatically compressed into `<session_summary>` (retaining decisions/facts/preferences), releasing context; LLM unavailable downgrades to retain original text. - **BM25 recall** — Each question automatically retrieves relevant historical conversations and facts as `<memory>` blocks injected into AI context (zero dependency, pure local, offline capable). - **User profile / facts (Layer A)** — Fades out stable facts (preferences, projects, identity), deduplicates accumulated confidence, distills into `profile.md`; viewable, forgettable. - **Degradable switch** — Environment variable `HESI_MEMORY_ENABLED=0` closes the subsystem, chat returns to original localStorage behavior, **zero behavioral change**. - **Old history migration** — First upgrade automatically imports browser old `localStorage['qcli-chat-history']` as the first session; can also use `node scripts/memory-migrate.js --file legacy.json` offline import. > Privacy: all memories are stored locally in `data/memory/`, not connected to any cloud; `facts.json` / `profile.md` can be manually edited or deleted. ### 🤝 AI Assistant × CLI Agent Collaborative Discussion (Round-Table) > This feature is Hesi's core collaboration scenario: let "AI assistants" and one or more "CLI Agents (like opencode) have **multi-round discussions** on the same issue, questioning, supplementing, and refining solutions. - **Multi-select discussion partners** — In the chat panel, click the "Discuss" button, multi-select (up to 4) from discovered CLI Agents as discussion objects - **Multi-round iteration** — Supports configuration discussion rounds, AI assistants and Agents take turns speaking, converging round by round - **Incremental extraction** — Each round only feeds Agent's **new output** to AI (preventing same text from filling up context) - **Favorites synchronization** — Discussion dropdown merges `/api/agents` and `/api/clis`, and reads left sidebar favorites, **favorite items default to checked, pinned, and marked with ★** - **Timeout and cleanup** — Single Agent session 5 minutes timeout automatic termination, completed sessions 5 minutes TTL after cleanup ### 🛡️ CLI Agent Rendering > Full-screen TUI CLI Agents (like **opencode**) draw ASCII interfaces/status bars in PTY, their rendering frames are literal text, stripped of escape sequences, polluting discussion text fed to AI (manifesting as "stuck in interface rendering, not providing substantial analysis"). - **Headless subcommand** — For Agents declaring headless mode, use non-interactive subcommands and **stdin pipeline** injection tasks (non-TTY, from the source to prevent TUI), outputting clean plain text - **Multi-CLI support** — `lib/cli-headless.js` `HEADLESS` maps current built-in four (all tested, task prompts unified stdin, no argv拼): `opencode` (`opencode run`), `claude` (`claude -p`), `codex` (`codex exec -`), `aider` (`--yes-always --no-auto-commits --no-pretty --no-stream`); add other CLI just need to make up a description - **Windows security points** — Headless execution on Windows starts with `shell:true`, argv will be re-parsed, so **multi-line/quoted prompts must be injected via stdin** (tested with `test/cli-headless.test.js` with malicious prompts covering `"`/newline/`rm -rf`) - **Fallback compatibility** — Agents not declaring headless still use PTY + escape cleaning (`lib/terminal-clean.js`), behavior unchanged - **TUI retention** — Human interactive terminal (`ws/agent.js`) and workflow (`ws/orchestrator.js`) TUI complete retention, unaffected ### 🌐 Browser Control (CDP) - **Automatic connection** — Detect Chrome/Edge CDP port (default `localhost:9222`) - **Page operation** — Navigation, forward/backward, refresh, screenshot, execute JS - **Element interaction** — Click, input, hover, scroll - **Console monitoring** — Real-time browser console log acquisition - **Tab management** — List/switch all browser tabs - **Browser farm** — Multiple isolated browser sessions parallel management - **User script** — Inject custom JS scripts in specified URL patterns automatically execute - **DOM Diff** — DOM snapshot comparison, tracking page changes - **Form auto-filling** — Automatic detection and filling of form fields - **Accessibility analysis** — Page accessibility issue detection - **Network monitoring** — Real-time capture of HTTP request/response, supporting HAR import/export ### 📄 Document Format Conversion - **AI-driven** — Through `convert_document` tool in AI dialogue direct conversion - **Multi-format support** — PDF / DOCX / PPTX / HTML / EPUB / LaTeX / RST / Markdown - **Pandoc-driven** — Automatic detection of system pandoc, supporting `PANDOC_PATH` environment variable - **Intelligent degradation** — Without pandoc, automatically use built-in Markdown→HTML converter ### 🔌 MCP Server - **Modular architecture** — `mcp/` directory contains tools, resources, security, session management submodules - **Session management** — `SessionManager` + `RingBuffer` + TTL automatic expiration - **Security layer** — Bearer authentication + audit logs + YAML policy file - **Cache layer** — LRU cache + METRIC statistics + heartbeat report - **AI bridging** — MCP → OpenAI Function Calling conversion - **Automatic restart** — Health check + exponential backoff restart - **Rate limiting** — Token Bucket algorithm, preventing tool call storm - **Output truncation** — Tool result automatic truncation (4K character limit) ### ✅ Code Quality & Security - **PTY environment variable filtering** — Automatic filtering of API_KEY/TOKEN/PASSWORD sensitive patterns - **Double-layer current limiting** — Global API current limiting + WebSocket message current limiting + upload current limiting + upload limiting - **Regression testing** — Terminal cleaning, discussion coordinator, stability (tool interruption/current limiting/stream completion/ring detection) three-layer verification - **Prettier + ESLint** — Code unified formatting + lint-staged pre-commit check### 🎙️ User Experience - **Command — Ctrl+K CLI and operation search **Voice Input/Output** — Web Speech API for input; TTS for reading AI responses - **Theme Customizer** — Built-in presets + custom saving - **Multilingual Interface** — English / Chinese for instant switching - **Notification System** — Toast notifications + Notification Center - **Custom CSS** — Real-time injection of custom styles - **Welcome Carousel** — Guided introduction for first-time## Quick Start ### Prerequisites | Item ||------|-------------| | **Node.js** | >= 0.0 | |** | >= 9.0.0 | | **** | Windows / mac Linux | ### Install ```bash git clone https:///qiuqiukofesi.git cd Hesi # (need to extract portable node to directory) npm install npm run build # Production build for frontend (outputs public/bundle.js) npx playwright install chromium # Optional, for browser control functionality cd Hesi/tray npm install ``` ### Start ```bash npm start # → http://localhost:3001 (default listen on 127.0.0.1 and ::1) npm run dev # Development mode () npm run # Start MCP service``` ### Environment Variables | Default Value | Description | |----------|---------------|-------------| | `PORT` | `3001` | HTTP service port | | `HOST` | `loopback` | Listen address; set0.0.0.0` will print warning | | `OPENAI_API_KEY` / `ANTHROPIC_API_KEY` | — | LLM provider keys (optional) | | `STABILITY_API_KEY` / `B_API_KEY` / `TAVILY_API_KEY` | — | Image/search keys (optional) | | `PANDOC_PATH` | — | Pandoc path (optional, for document conversion) | | `QCLI_ACCESS_TOKEN` | `""` | Access token; set to require authentication forapi` and WebSocket (loopback exempt by default) | | `QCLI_TOKEN_REQUIRE_LOOPBACK` | `""` | Set to `1` to enforce token even on loopback | | `QCLI_CORS_ORIGINS` | `""` | Comma-separated CORS whitelist | | `QCLI_POLICY_PATH` | `""` | MCP security policy file (default `blocklist`) | | `QCLI_WITH_MCP` / `QCLI_MCP_TOKEN` | `""` | Auto-start MCP / MCP Bearer token | | `QCLI_AUDIT_LOG` | `""` | MCP audit log path | | `QCLI_SESSION_TTL` | `900000` | Session idle expiration (ms) | | `QCLI_MAX_SESSIONS` | `10` | Maximum concurrent sessions | > 💡 See complete variables in `./.env.example` ## Desktop Tray Bundle Hesi also provides an **out-of-the-box offline intelligent package**: comes with a portable Node.js, no installation required, and can be started by double-clicking. ```text Windows: Double-click tray.bat macOS : Run ./tray.sh in terminal (need chmod +x tray.sh for first time) Linux : Run ./tray.sh in terminal ``` - After starting, Hesi icon appears in the tray, and automatically opens `http://127.0.0.1:4264` - On the welcome page, click 「AI Agent (one-click install)」 to install OpenCode / Codex offline - Tray menu: Open Hesi / Open (CDP mode) / Stop service / Exit - By default, only binds to the local loopback address and does not expose to the LAN - macOS 「Cannot verify developer」: `xattr -dr com.apple.quarantine .` and retry - Port occupied: set `PORT` environment variable to change port ## Architecture ``` ├── server.js # Express entry + static file service ├── ws-handler.js # WebSocket connection management + PTY (core terminal logic) ├── cli-discovery.js # CLI auto-discovery engine (including disk registry cli-registry.json) ├── cli-registry.json # CLI registry (agent list / favorite source, runtime persistence) ├── preset-loader.js # Preset loader ├── rate-limiter.js # API rate limiting ├── ring-buffer.js # Ring buffer ├── mcp-server.js # MCP sidecar entry ├── mcp/ # Modular MCP architecture (tools/resources/security/session) ├── ws/ # WebSocket subsystem │ ├── pty.js # PTY creation abstraction + createHeadlessExec (headless Agent execution) │ ├── pty-policy.js # PTY policy engine │ ├── message-dispatch.js # WebSocket message routing │ ├── agent.js # Human-computer interaction Agent terminal (保留 TUI) │ ├── orchestrator.js # Workflow orchestration TUI, support single ws concurrent multi-workflow) │-employee.js # Digital employee team (role/persona/task assignment) │ ├── digital-employee-worker.js # Digital employee task executor (reuse agentPool, real task execution) │ └── context-store.js # Shared context storage / # RESTful API routes │ ├── # AI chat + discuss (AI × CLI Agent roundtable) │ ├── ai-tools/ # Agent pool-pool.js) + synchronous delegation (builtin/agent.js) ├── clis.js / agents.js # CLI / Agentincluding category, support favorite synchronization) │ └── ... # Other routes ├── lib/ │ ├── cli-headless.js # Headless Agent description table (opencode/claude/codex/aider, all stdin injection) │ ├── asset-hash.js # bundle.js/lazy-bundle.js content hash (?v= cache penetration) │ ├── terminal-clean.js # TUI escape cleaning (CSI/OSC/raw ESC stream cleaning) │ ├── env-filter.js # PTY environment variable filtering │ ├── access-auth.js # Optional access token authentication │ └── mcp-process.js # MCP subprocess management ├── public/ # resources (bundle.js esbuild) ├── cli-presets/ # CLI preset configurations ├── workflows/ # Preset workflow orchestration ├── plugins/ # Plugin system ├── tray/ + tray.bat/tray.sh # Desktop tray launcher (offline portable package) └── node/ # Portable Node.js runtime offline package) ``` ## Tech Stack | Layer | Technology | |----|------------| | **Backend** | Node.js | | **Terminal** | node-pty + xterm.js + WebGL | | **Communication** | WebSocket ( + SSE | | **Build** | esbuild | | **AI** | OpenAI API / Anthropic API | | **Browser Control** | Playwright + CDP | | **Storage** | IndexedDB + safeStorage | | **Code Quality** | ESLint + Prettier + Husky | | **Testing** | node --test + plans/ regression suite | | **Documentation** | OpenAPI 3.0 + JSDoc | | **Speech** | Web Speech API | | **Charts** | Self-developed Canvas engine (ChartCore) | ## Testing Regression suite located in `plans/` (pure Node script, no framework): ```bash node plans/verify-terminal-clean.js # Terminal escape cleaning (9 items) node plans/test-discuss.js # Discussion coordinator module contract regression (7 items: load/export/shape/async/arity/cache) node plans/test-stability.js # Stability37 items: tool interruption/rate limiting/stream completion/loop detection) ``` Syntax and structure check: ```bash npm run check:server # All server-side modules node --check syntax check npm run lint # ESLint ``` ## Scripts | Command | Description | |----------|-------------| | `npm start` | Start production service | | `npm run dev` | Development mode (`--watch` hot reload) | | `npm run build` | Build frontend (esbuild compression, output public/bundle.js| `npm run build:dev` | Development build (with sourcemap) | | `npm run watch` | Frontend listen mode | | `npm run mcp` | Start MCP service independently | | `npm run check:server` | Server-side module syntax check | | `npm run lint` / `npm run format` | ESLint / Prettier | ## Contributing Contributions welcome! Feature requests, bug reports, and code PRs. Complete guide (including build pitfalls, husky hooks, gh-pages process, bundle dual-track trap) see **[CONTRIBUTING.md](./CONTRIBUTING.md)**. ### Development Process ```bash git clone <your-fork> cd Hesi npm install npm run dev npm run build ``` ### Code Style - Backend: CommonJS (`require`/`module.exports`) - Frontend: ESM (`import`/`export`), packaged by esbuild - New features please attach regression tests (located in `plans/`) ## Secure Deployment Hesi is essentially a **local-first** terminal/browser hub: it executes arbitrary commands via WebSocket and controls the browser via CDP. Therefore, the default configuration follows the principle of "minimum exposure" and is secure out of the box. ### About `npm audit` warnings After `npm install`, `npm audit` may report some vulnerabilities (including a moderate one from `@hono/node-server`). These come from **indirect dependencies of `@modelcontextprotocol/sdk`**, Hesi itself uses the **Express** server, **does not call Hono's `serve-static`**, and the service **only listens to the local loopback**, not exposing it to the public network. Therefore, this path is dead code not called, **has no actual impact on local deployment**, and is a false positive in the supply chain, no need to handle. The rest of the vulnerabilities that can be directly fixed have been upgraded to compatible versions via `npm audit fix` (not `--force`). ### Default Security Posture (Secure Out of the Box) | Item | Default Behavior | |----|----------| | Listen Address | address `127.0.0.1` and `::1` (local only). Set=0.0.0.0` will warning | | CORS | Only allow same-origin/loopback; cross-domain requires explicit whitelist with `QCLI_CORS_ORIGINS` | | Access Token | Disabled when `QCLI_ACCESS_TOKEN` is not set; set to require token for all sensitive `/api` and WebSocket (loopback exempt) | | Command Policy | `blocklist` mode,-in dangerous command blacklist be overridden by `QCLI_POLICY_PATH` | Limiting | Global API + WebSocket message + upload rate **local loopback exempt by default** | Isolation | node-pty degrades gracefully if not compiled, PTY functionality | | Upload Directory | User uploads `uploads/.user/` (hidden, authenticated and | ### Public Network / Multi-Person Deployment Checklist ⚠️ Only expose to non-local network when all conditions are met. 1. ** Token**: `QCLI_ACCESS_TOKEN=<strong-random-token>`, HTTP header `Authorization: Bearer <token>`, WebSocket append `?token=<token>` 2. **Tighten CORS**: `QCLI_CORS_ORIGINS=https://your-frontend.example.com` 3. **Strengthen Command Policy**: Use `QCLI_POLICY_PATH` to point to `blocklist`/`allowlist` policy, or tighten to `allowlist` 4. **MCP Authentication**: `QCLI_MCP_TOKEN` + keep `QCLI_AUDIT_LOG` enabled 5. **Reverse Proxy**: Place Nginx/Caddy in front to enable HTTPS/HSTS, restrict `/api/uploads` origin 6. **Do not run as root**, and update dependencies regularly ### Security Baseline Self-Check ```bash [ "$HOST" = "0.0.0.0" ] && echo "WARN: HOST=0.0.0.0 exposes all interfaces" || echo "OK: loopback by default" [ -n "$QCLI_ACCESS_TOKEN" ] && echo "OK: access token set" || echo "WARN: no access token" ``` ## License MIT License — see [LICENSE](./LICENSE) for details. <p align="center"> <sub>Built with ❤️ by Hesi Contributors</sub> </p> ## Why Hesi An open-source platform for running any CLI/terminal in the browser, core values: - **🌐 Browser Terminal** — `node-pty` + `xterm.js` + WebSocket, native terminal experience, multi-tab independent operation, session persistence, no need for SSH. - **🤝 AI Agent Collaboration** — Let AI assistants and opencode/codex and other CLI Agents have multi-round "roundtable discussions", and headless execution from the source eliminates TUI pollution🔌 Ecosystem** — Built-in enterprise wechat, Feishu and other, a hub connecting and business system. - **💾 Offline Portable Comes with a portable Node.js runtime desktop package, double-click `tray.exe use, no cloud. 📖 Open Source MIT license, local can be privatized, capabilities auditable-hosted. ## > Replace the badges with actual CI/version/download links. [ [](https:///qiuqiukoss/Hesi/releases](https:///badge/license-blue.svg)](./LICENSE) [](https://github.com/qiuqiukof-oss/Hesi) [](https://github.com/qiuqiukof-oss/Hesi/stargazers) ## Demo - 3 demo: [docs](./docs/demo-script.md) (out of the box → AI roundtable → terminal/browser control → → offline portable). ready**: Desktop double-click `tray.exe` (Windows) `tray.sh` (macOS/Linux) to start, comes with portable Node, default only bind local loopback, no installation required, no internet connection needed the "Desktop Tray Bundle (Offline Portable Package for details. ## Contributing Contributions welcome requests, bug reports, documentation, and code PRs are all encouraged. Complete contributor guide (development process, code style, build traps, husky hooks, gh-pages release process) see **[CONTRIBUTING.md](./CONTRIBUTING.md)**. ### Reference Documentation - Getting started guide for newcomers: [docs/getting-started.md](./docs/getting-started.md) - Architecture whitepaper: [docs/architecture.md](./docs/architecture- Compliance and-inspection: [COMPLIANCE.mdPLIANCE.md) - Secure deployment guide:_DEPLOY.md](./SECURE_DEPLOY