A free, open-source 1:1 alternative to Grok Bot and Muse. Message Claude Code, Codex, Cursor and 40+ coding agents as persistent bots from your iPhone, Mac or Linux desktop. Group chats, threads, appr
Baiqiao Zhang is a PhD student in Computer Science and Engineering at HKUST, advised by Prof. Xiaojuan Ma, working on human-AI alignment between developers and AI coding agents.
Taiwan-built, self-hosted AI agent harness in a single Go binary — writes its own tools, sandbox-tests them, and shares them over MCP with Claude Code and Codex.
A powerful, cross-platform AI-enhanced terminal built with Tauri 2, React 19, and Rust. AITerm seamlessly integrates the traditional command-line experience with advanced AI capabilities, autonomous agents, a task board that dispatches work to Claude Code on your behalf, Agent Collaboration that organizes several AI engineers into a team that builds, reviews, merges and verifies a project, and project requirement management.
Beyond cloud providers, AITerm connects to local models (Ollama, LM Studio, oMLX, and other OpenAI-compatible endpoints) so inference runs on your own machine — no API cost and nothing sent off-box — which pays off for high-volume automation. It can also bridge the Claude Code CLI to any provider you have configured, so you run Claude Code on your OpenAI, Gemini, or local models.
✨ Key Features
Integrated AI Providers: Out-of-the-box support for multiple LLM providers including OpenAI, Anthropic, Google, Ollama (local inference), and OpenAI-compatible endpoints. You can also sign in with a subscription account — Claude, ChatGPT, Gemini, or GitHub Copilot — with no API key to obtain.
AI Command Flow (/ai): Type /ai <query> directly in the terminal to generate and preview commands with risk-level assessment before execution.
Local Models · Save on Tokens: Connect to Ollama, LM Studio, oMLX, and other local / OpenAI-compatible endpoints — inference runs on your own machine with no API cost and nothing sent off-box. Or use the Claude Code Bridge to run the Claude Code CLI on any provider you have configured.
Structured Command Blocks: Every command and its output render as a distinct card instead of raw scrollback. The live terminal pane auto-expands only while a command is running and shrinks back down when idle, keeping the screen uncluttered.
Autonomous Agent Loop (/agent): Multi-step, goal-driven agentic execution for complex tasks, with built-in guardrails and fallback to manual confirmation for dangerous operations.
LoopStudio (Loop Engineering): A visual multi-agent orchestration workbench — an orchestrator breaks a goal down and dispatches sub-agents to work in parallel, with safety gates and a live execution trace, running autonomously until the goal is verified done instead of requiring back-and-forth confirmation.
Task Board: A kanban board that hands work to Claude Code for you. Write what you want done as cards — title, which folder to work in, detailed instructions, attachments if needed — and drag one to Queued: AITerm opens a tab, launches Claude Code, delivers your instructions, and moves the card to Done with a success or failure outcome when it finishes. Queue up an evening's work instead of driving each task by hand. Cards live in projects (self-contained folders you can copy to move or share), several projects open as tabs at once, and a global concurrency cap keeps the machine sane. Interactive mode marks work you intend to discuss with Claude Code yourself, exempt from the cap and from stall detection. Finished work keeps a clean transcript you can reread, and "Requeue" clones a card to run the same thing again. Search filters all four columns by title, body or folder. Each card can run on Claude Code or OpenAI's Codex CLI.
AI Work Report: Turn a project's board into a formatted document. Pick a style, choose a model, and AITerm reads every card — including the full transcripts of finished work — and writes an HTML report that is saved into the project folder and kept as history. A finished card's summary is written once and cached, so a second report only processes what is new.
🏗️ Architecture
AITerm uses a modern Desktop architecture communicating via Tauri IPC (invoke + events):
Frontend (src/): React 19 shell, xterm.js terminal views, component-local state management, and an intelligent frontend router for handling AI commands and streams.
Backend (src-tauri/src/): Rust-powered high-performance core handling PTY lifecycles, AI provider routing, SQLx-backed database connections (including a Java JDBC sidecar for DB2), Telegram bot integration, and secure local configuration (OS keyring).
🚀 Getting Started
Prerequisites
Rust 1.78+ (rustup show)
Node.js 20+ (node -v)
Windows only: Windows 11 with WebView2 runtime and MSVC build tools.
Development
npm install
npm run tauri:dev
Note: The first build will take a few minutes as Rust dependencies are compiled. Subsequent runs are much faster.
Testing
# Frontend tests (Vitest + React Testing Library)
npm run test
# Rust unit & integration tests
# --workspace is required: src-tauri/Cargo.toml is both a package and the
# workspace root, so a bare `cargo test` runs only the `app` package and
# silently skips aiterm-core and aiterm-host.
cd src-tauri && cargo test --workspace
# Type checking
# NOT `tsc --noEmit`: the root tsconfig.json is a solution file ("files": []),
# so it checks nothing and always exits 0.
npx tsc -b
🖥️ aiterm-host (remote host)
Run a shell on a machine with no desktop — a cloud VM, a container, a NAS —
and connect to it from AITerm to drive it with AI. The AI runs on your
desktop, so the host needs no AI configuration or API keys.
Install
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/jamesju9999/AITERM/master/scripts/install.sh | sh
# Homebrew (trust the tap first — see below)
brew trust jamesju9999/tap
brew install jamesju9999/tap/aiterm-host
# npm (or run without installing via npx)
npx aiterm-host --print-connection
# Container
docker run -it --rm -p 8022:8022 -e AITERM_HOST_KEY=<your-key> \
ghcr.io/jamesju9999/aiterm-host:latest
# Windows
irm https://raw.githubusercontent.com/jamesju9999/AITERM/master/scripts/install.ps1 | iex
Homebrew: invalid syntax in tap! is not a syntax error. Recent Homebrew
refuses to tap a third-party repository you have not trusted, and the message
it prints — Cannot tap jamesju9999/tap: invalid syntax in tap! — points at
the formula rather than at the trust check. Run brew trust jamesju9999/tap
once and the install proceeds.
The install scripts verify every download against the release's SHA256SUMS
and abort on a mismatch — there is no "install anyway" option.
aiterm-host has its own version numbers and release notes
(CHANGELOG-host.md), published from host-v* tags
independently of the desktop app.
Use
aiterm-host --bind 0.0.0.0 --port 8022
It prints an address, a port and a key. In AITerm, open Connect to a remote
terminal, expand the manual-address section, and enter the address, port and
key (leave the 6-digit code blank).
The key lives in ~/.config/aiterm-host/key and survives restarts, so a saved
connection keeps working. Across networks, use an address the viewer can reach
(Tailscale, a VPN, or an SSH tunnel with --bind 127.0.0.1).
一個強大且跨平台的 AI 智慧終端機，基於 Tauri 2、React 19 和 Rust 打造。AITerm 將傳統的命令列體驗與先進的 AI 能力、自主代理（Agent）、可自動把工作派給 Claude Code 執行的工作看板、讓好幾個 AI 工程師分工開發、互相審查、合併並實際驗收的 Agent 協作，以及專案需求管理完美結合。
除了雲端供應商，AITerm 也能串接本地端模型（Ollama、LM Studio、oMLX 及其他 OpenAI 相容端點），推論在自己機器上跑——不花 API 費用、資料不外送——量大的自動化任務尤其划算。也可以用 Claude Code 橋接，讓 Claude Code CLI 跑在你已設定的任何供應商上（OpenAI、Gemini 或本地模型皆可）。
✨ 核心功能
多重 AI 供應商整合：開箱即支援 OpenAI、Anthropic、Google、Ollama（本地端推論）以及 OpenAI 相容端點等多種語言模型。也可以直接用訂閱帳號登入——Claude、ChatGPT、Gemini、GitHub Copilot 皆可，不必另外申請 API Key。
AI 指令工作流 (/ai)：直接在終端機輸入 /ai <查詢> 即可生成指令，並在執行前提供風險等級評估與預覽。
本地模型・省 token 費用：接 Ollama、LM Studio、oMLX 這類本地 / OpenAI 相容端點——推論在自己機器上跑，不花 API 費用、資料不外送。或用 Claude Code 橋接，讓 Claude Code CLI 跑在你手邊任何一個供應商上。
工作看板：把工作排進去，AITerm 自動派給 Claude Code 執行。把想做的事寫成卡片（標題、要在哪個資料夾做、詳細指示，需要的話還能附上檔案），拖到「待執行」，AITerm 就會開一個分頁、啟動 Claude Code、把指示送過去，做完自動搬到「已完成」並標上成功或失敗——一次排好幾件晚上再看結果，不必守在電腦前一件一件下指令。卡片歸屬於專案（自成一體的磁碟資料夾，複製即可搬移或分享），可同時開多個專案分頁，並行上限為全域設定。「互動模式」用於你打算自己跟 Claude Code 討論的工作，不佔並行額度也不會被判定卡住。完成的工作留有乾淨的對話記錄可回顧，「重新派工」會複製一張新卡片重跑同一件事。搜尋可依標題、內容或工作目錄即時過濾四個欄位。每張卡片可以選擇用 Claude Code 或 OpenAI 的 Codex CLI 執行。
AI 工作報告：把一個專案的看板整理成一份文件。選風格、選模型，AITerm 會讀過每一張卡片（含已完成工作的完整對話記錄），產出排版好的 HTML 報告，存進專案資料夾累積成歷史。已完成的工作摘要只整理一次就快取起來，第二次產報告只需處理新完成的部分。
Archive: Finished cards would otherwise pile up forever. Archive them one at a time or a whole column at once and the board keeps only what is in flight. Nothing is deleted — cards, transcripts and attachments stay put, the archive view searches and pages through them, and any card can be restored to the board.
AI Refine: Jot a rough note into a task's body, press Refine, and the AI rewrites it into a brief Claude Code can execute — goal, scope, constraints, acceptance criteria — without inventing work you did not ask for. One click restores your original.
Agent Collaboration: Turn several Claude Code / Codex agents into a team on a canvas. Engineers (as many as you like) each work on their own git worktree and branch, so they truly run in parallel — drag a module onto an engineer and it starts. A reviewer reads every finished module's diff and passes it or sends it back with reasons (up to 2 rounds, then you decide); on a pass it writes a hand-off note on what the module provides. An integrator merges every branch once all modules pass, and an optional verifier builds, runs and actually uses the merged result — clicking through a browser for web apps — checking each acceptance criterion with screenshot evidence before your repository is fast-forwarded. Every agent starts with a shared big picture: a project brief you write once (with AI refine), plus an auto-generated team map of who owns what, progress, dependencies and hand-off notes. Agents ask for decisions in a Decisions inbox (with desktop and Telegram notifications), wait out usage limits and resume, and pick up the same conversation after an app restart. Also: spec-first planning, task dependencies, per-role models, per-project token caps, team templates, and a guided tour. Requires the Claude Code CLI (Codex optional).
Code Assistant: A dedicated tab where the AI answers questions about any project directory — it scans the file tree, searches and reads source files on its own, can render Mermaid diagrams, and lets you export the conversation to Markdown.
Multi-turn Chat Sidebar: Persistent, context-aware AI chat directly beside your terminal for troubleshooting, code generation, and brainstorming.
AI Document & Chart Panel: Reports, comparison tables, and charts the AI produces open in a formatted panel beside the conversation instead of raw markup — bar/line/pie charts render directly, switch to a table view for exact numbers, and export to an HTML file in one click. Available across the terminal and remote AI panels, Code Assistant, Knowledge Base, and database / cross-database Q&A.
Knowledge Base: Create a notebook pointed at a folder and AITerm chunks, embeds, and indexes every document inside it, so you can ask natural-language questions with answers that cite the source documents. Works with Ollama or any OpenAI-compatible embedding provider.
Multi-Database Connections: Native support for PostgreSQL, MySQL, SQLite, and ODBC. DB2 works on both macOS and Windows via a built-in Java JDBC sidecar — no driver installation required.
File Explorer: A built-in sidebar file browser kept in two-way sync with the terminal's working directory — navigate in the terminal and the explorer follows, or click the "cd" button (or the folder picker built into the command input) to jump the terminal to any folder instantly.
Command Search & Bookmarks: Ctrl+F searches across the entire block history with live scroll-to-match; Ctrl+Shift+R opens a bookmarks picker to save and re-fill frequently used commands.
Multilingual UI: Full support for English and Traditional Chinese, switchable at runtime with no restart required.
Remote Terminal AI Panel: When connected to a colleague's shared terminal, Ask AI is the same full chat panel as local — free-form Q&A, provider switching, and browsing or continuing past conversations, all through your own configured provider so the other machine needs no AI setup; /agent works in a remote tab too.
Telegram Remote Control: Send commands to your terminal from your phone via a Telegram bot, with results streamed back to the same chat.
Requirement Management System: AI-assisted discussion for feature planning, automatically generating and saving specifications (SDD) into structured, project-managed directories.
MCP Tools: Connect any MCP (Model Context Protocol) server and the AI can read and write files, query data, and call external services directly — no terminal command in between. Every tool call and its result stays visible and expandable.
Claude Code Bridge: Point the Claude Code CLI at any provider you have configured in AITerm — run Claude Code on your OpenAI, Gemini, or local models, and switch between them at any time.
Next-Step Suggestions: While Claude Code, Codex, Gemini CLI or a similar AI CLI runs in the terminal, a side panel reads the screen and suggests 3–5 prompts to send next — click to fill, double-click (or ➤) to send. Set a big goal per tab and let the AI break it into milestones; suggestions steer toward the next unfinished one. A prompt helper rewrites rough requests, and likely API keys, tokens and passwords are masked before the screen is sent.
Computer-Use Hub: A global hotkey (default Ctrl+` ) summons or hides AITerm from anywhere, a tray icon keeps it running in the background, and it can start at login. With Computer Use enabled, the AI (in Ask AI or any external MCP client) can list apps, read UI elements, take screenshots, click, type and scroll in other applications — asking before it touches each app for the first time, re-confirming destructive clicks, and logging every action.
OpenAI-Compatible API Server: Let Cursor, Cline, Open WebUI or your own scripts talk to AITerm and use any provider you have configured in it — subscription sign-ins, API keys or local models. Requests are routed through a model alias table, need an API key, support streaming and tool calling, and accept LAN connections only if you opt in.
System Integration: Use AITerm as the system terminal on macOS and Linux; right-click a folder → "Open in AITerm" in macOS Finder and Windows Explorer; on Windows, run commands that need administrator rights elevated in the same tab. AITerm remembers window size and position and warns before closing while work is running.
Command History Hints: As you type, a grey hint suggests a command you have typed before; press Tab (configurable) to accept it.
Robust Terminal Engine: Powered by xterm.js and portable-pty for a fast, native-feeling shell experience across Windows, macOS, and Linux.
Claude Code 橋接：把 Claude Code CLI 接到 AITerm 已設定的任何供應商——用你手邊的 OpenAI、Gemini 或本地模型跑 Claude Code，模型可隨時切換。
下一步建議側欄：在終端機跑 Claude Code、Codex、Gemini CLI 這類 AI 命令列工具時，側欄會看目前畫面，提出 3 到 5 個下一步可以送出的提示詞——單擊填入、雙擊（或按 ➤）直接送出。可以替每個分頁設定大目標，讓 AI 拆成里程碑，建議會朝第一個未完成的里程碑推進；提示詞助手能把粗略的需求改寫清楚；送出畫面前會先遮罩疑似的 API 金鑰、token 與密碼。
電腦操作入口：全域快捷鍵（預設 Ctrl+` ）在任何地方叫出或隱藏 AITerm，托盤圖示讓它常駐背景，也可以開機自動啟動。開啟「電腦操作」後，AI（Ask AI 或任何外部 MCP 用戶端）能列出應用程式、讀取畫面元件、截圖、點擊、打字、捲動其他應用程式——每個應用程式第一次操作都會先問你，刪除／送出類的按鈕每次都再確認，所有操作都有紀錄。
OpenAI 相容 API 伺服器：讓 Cursor、Cline、Open WebUI 或你自己的腳本直接連到 AITerm，使用你在 AITerm 設定好的任何供應商（訂閱登入、API 金鑰、本機模型皆可）。模型名稱依別名表轉送，請求要帶 API key，支援串流與工具呼叫，要自己勾選才開放區網連線。
系統整合：macOS 與 Linux 可把 AITerm 當成系統的終端機；在 macOS Finder 與 Windows 檔案總管對資料夾按右鍵「在 AITerm 開啟」；Windows 需要系統管理員權限的指令可以直接在同一個分頁提權執行。AITerm 會記住視窗大小與位置，有工作在進行時關閉會先提醒。
指令歷史提示：輸入時用灰字提示你之前打過的指令，按 Tab（可改）補上。
強大的終端機引擎：基於 xterm.js 與 portable-pty，在 Windows、macOS 與 Linux 上提供快速、原生的 Shell 體驗。
