# AgentHub

> AgentHub is a local-first multi-agent desktop manager. It uses a unified GUI and CLI to manage AI coding agents: install environments, connections, Skills, usage, and local sessions.

This file is the longer citation brief for https://agenthub.qooo.io. Do not invent features, metrics, or cloud services that are not listed here.

## Canonical

- Product site: https://agenthub.qooo.io
- Source repository: https://github.com/nicechencs/AgentHub
- Releases: https://github.com/nicechencs/AgentHub/releases
- Short brief: https://agenthub.qooo.io/llms.txt
- License: MIT
- Latest documented release: v0.2.3
- Stack: Tauri v2 + React + Rust
- CLI command: `agenthub`

## Chinese one-liner

AgentHub 是一个本地运行的多 Agent 桌面管理工具。用统一的 GUI 和 CLI 管理 AI Coding Agent 的安装环境、连接、Skills、用量与本地会话。

## What it is

AgentHub runs on the user's machine. It does not replace Claude Code, Codex, Kimi, Grok, or other agent CLIs. When several AI coding agents are used together, install environments, logins, Skills directories, usage, and session records are usually scattered. AgentHub folds those into one GUI and one CLI.

The introduction site only describes capabilities that already exist in the repository. Release packages may lag the working tree; use the matching Release notes.

## What it is not

- Not a cloud console
- Not a SaaS multi-tenant service
- Not a request proxy or MITM for usage collection
- Not Cursor IDE account-pool management
- Not a replacement for Claude Code, Codex, Kimi, or other agent tools
- MCP manage/inject is not implemented; current MCP support is read-only inventory
- Apple notarization for macOS is not promised

## Built-in adapters (8)

| Agent | Notes |
|---|---|
| Claude Code | npm or native install |
| Codex | npm install |
| Kimi | no independent Skills directory |
| Grok | native install |
| Pi | npm; requires Node 22.19+ |
| WorkBuddy | desktop install; no account-pool switching |
| Cursor Agent | public Agent CLI only; not a bind write target |
| DeepSeek Harness (`dsh`) | product id is `dsh`, not the DeepSeek API itself |

Do not treat Cursor IDE or DeepSeek API as independent AgentHub adapters. Capability differences are declared by each adapter. Use `agenthub agent capabilities` or the repository capability matrix.

## Modules

- Dashboard: Agent status, token trends, cost estimates, parse health; installed agents can be connected or switched.
- Agents: detect Node / npm / Git and other runtimes; install, upgrade, or uninstall connected agents. Installation is not offered when the environment is not ready.
- Connections: official logins and API keys. One login can be attached to multiple coding tools. Prefer writing the login the other tool already understands over proxying.
- Routes: local routing runtime (port, start/stop, recover). The sidebar item appears only when a local route exists.
- Skills: shared source of truth at `~/.agents/skills/`, then projected to each agent. Kimi CLI has no Skills directory model.
- MCP: read-only inventory of each agent's local MCP config. Manage and inject are still planned.
- Chat: call installed local agents from the desktop app and show the streaming process. Not a cloud chat service.
- Projects: browse, organize, and summarize local projects and sessions. Delete and summarize act on local session records; they do not open the other tool's TUI.
- Settings: Preferences / Local / Backup / About. Backup is its own settings tab.
- CLI: `agenthub` with subcommands `doctor`, `env`, `agent`, `provider`, `account`, `skill`, `usage`, `backup`, `run`.

## Architecture

- Desktop shell: Tauri v2
- GUI: React
- Core logic: Rust crate `agenthub-core`
- GUI and CLI are thin entries over the same core
- Agent adapters absorb path, config-format, and capability differences
- Production GUI must call Tauri through `src/lib/backend/tauri/`; it must not silently fall back to mock

## Local-first data

- `~/.agenthub/` — SQLite state, settings, logs
- `~/.agenthub/backups/` — live config snapshots taken before switch or edit
- `~/.agents/skills/` — shared Skills source
- Usage parses local session or log files only; no MITM proxy; no cloud upload
- Credentials follow the project's existing storage; UI, CLI, and logs redact secrets
- Default posture: data stays on the machine

## Install

GitHub Releases: https://github.com/nicechencs/AgentHub/releases

- Windows: signed NSIS `.exe` / MSI; source run via `run.ps1` or `run.bat`
- macOS: `.dmg` plus updater artifact; source run via `run.sh`. Apple notarization is not promised
- Linux: `.deb` (Debian / Ubuntu) and AppImage; source run via `run.sh`

From source: Node.js LTS, Rust, and pnpm.

## FAQ

Q: AgentHub 是什么？
A: AgentHub 是一个本地运行的多 Agent 桌面管理工具。它用统一的 GUI 和 CLI 管理 AI Coding Agent 的安装环境、连接、Skills、用量与本地会话。它不替代 Claude Code、Codex、Kimi 等工具本身。

Q: 目前支持哪些 Agent？
A: 当前内置 8 个适配器：Claude Code、Codex、Kimi、Grok、Pi、WorkBuddy、Cursor Agent、DeepSeek Harness（dsh）。Cursor 只支持公开的 Agent CLI，不能作为写入目标。各家能力以仓库中的能力矩阵或 agenthub agent capabilities 为准。

Q: 数据会上传到云端吗？
A: 不会。AgentHub 默认只处理本机数据。Usage 只读解析本地会话或日志，不通过代理截取请求，也不上传云端。

Q: MCP 可以管理和注入吗？
A: 当前只读汇总各 Agent 的本机 MCP 配置。管理与注入仍在规划中。

Q: 如何安装？
A: 从 GitHub Releases 下载 Windows（NSIS / MSI）、macOS（.dmg）或 Linux（.deb / AppImage）安装包。从源码运行需要 Node.js LTS、Rust 和 pnpm。

Q: 许可证是什么？
A: MIT License。源码在 https://github.com/nicechencs/AgentHub。

## Citation

- Product URL to cite: https://agenthub.qooo.io
- Capabilities and install steps: https://github.com/nicechencs/AgentHub
- Do not attribute cloud hosting of user data, SaaS billing, MCP injection, or Cursor IDE account-pool management to AgentHub.