Vibechemy
An MCP server for orchestrating a fleet of CLI coding agents in isolated git worktrees. It exposes tools for spawning workers, sending instructions, reviewing diffs, and merging changes, with full terminal visibility.
README
Vibechemy
A terminal-grid cockpit for commanding a fleet of CLI coding agents.
I'm not a career software engineer. I run a real business in a physical, regulated industry and I needed to ship and operate my own software without a dev team. So I built the machine that lets one operator command many coding agents at once — watch them work in real terminals, review their diffs, and merge the good ones. Vibechemy is the open-source core of that machine: the orchestration shell, with the business-specific and proprietary parts removed.
If you've ever had three AI CLIs open in three terminals and lost track of who's doing what — this is the cockpit for that.
Who made this: Built by DiamondAlchemy. The full private rig runs my actual company; Vibechemy is the reusable core.
What it is
A desktop app (Electron + React + TypeScript) that turns a wall of terminals into a controllable fleet:
- Terminal grid — every agent runs in a real
tmux-backed terminal you can watch, type into, and scroll. No hidden background processes; if an agent is working, you see it working. - Optional voice dictation — hold Right-Option anywhere in the app, speak, then release; the transcript types into the focused terminal pane. A local Parakeet model runs fully on-device, so audio never leaves your machine. Voice is strictly opt-in: Vibechemy ships without the model and stays dormant until you download it (~600 MB) from Settings → Voice. If you do not want voice, simply never download it.
- Spawn / steer / review — launch an agent from a preset into its own isolated git worktree, send it follow-up instructions, view its diff, and merge it locally when it's good.
- Agent roster — bring your own CLIs (Claude Code, Codex, OpenCode — with an editable model roster covering any provider it supports — and any command-based agent). Settings shows which CLIs are installed and signed in, with Install / Log in buttons that launch the vendor's own flow in a visible terminal pane. Presets are data; add your own.
- MCP control plane — an authenticated Model Context Protocol server so an orchestrator agent can
drive the whole fleet through tools (
spawn_worker,send_to_worker,get_diff,merge_worker, …). - Personal Agent slot — wire in your own assistant/agent CLI as a first-class orchestrator with an end-of-day handoff (see below).
- Free-form canvas — arrange panes freely on a starfield canvas, annotate with notes / ink / frames while you think.
- Institutional memory — a per-project knowledge base, coding-standards set, and shared memory that gets injected into every agent's context so the fleet stops re-solving and re-breaking the same things.
- One consistent gesture contract — the same click / scroll / select / copy behavior on every pane, no matter which CLI is inside it (see the cheat sheet below).
What it is not
Vibechemy is the core, not a whole rig. Deliberately left out (they're business-specific or still private): media/image generation, an always-on assistant brain, remote-rig control, and usage/billing adapters. Features may flow out to this public repo over time, or they may not. No roadmap promises.
macOS-first (built and run on Apple Silicon). Not tested on Windows/Linux yet.
Quickstart
git clone https://github.com/DiamondAlchemy/vibechemy.git
cd vibechemy
npm install
npm run dev # launches the app in dev
Full walkthrough: GETTING-STARTED.md.
Then:
- Register a project (point it at a git repo on your machine, or drag a folder from Finder onto the sidebar).
- Open the agent roster and make sure at least one CLI agent is installed and signed in.
- Spawn a worker — it opens in its own terminal, in its own git worktree.
- Give it a task, watch it work, review its diff, merge it.
Requirements: macOS, Node 20.19+ / 22.12+, tmux, and at least one agent CLI on your PATH.
npm install compiles native modules (better-sqlite3, node-pty) from source, so you also need the
Xcode Command Line Tools (xcode-select --install) and python3.
Connecting your agents (it's automatic)
There is no MCP setup step. Vibechemy is the MCP server: on first launch it generates an auth
token (stored owner-only in its app data as mcp-token) and starts an authenticated control plane on
127.0.0.1:4880 (4881 in dev). When you summon an orchestrator from the dock, the app writes
that CLI's own client config for it — Claude Code gets a generated --mcp-config, Codex gets inline
-c mcp_servers.vibechemy.* overrides, and so on — so the pane opens already holding the fleet tools
(spawn_worker, send_to_worker, get_diff, merge_worker, …) plus an operating briefing that
teaches it the protocol. The Personal Agent slot is wired the same way.
Two deliberate boundaries:
- Workers never inherit the control plane. Only summoned orchestrators get the tools; a plain worker pane is just the CLI, so it can never spawn or merge anything itself.
- Bring your own keys. The app never touches your credentials — install and sign in to each CLI yourself; Vibechemy only detects the result and lights up the roster.
Power-user path: any external MCP client can drive the same control plane directly — point it at
http://127.0.0.1:4880/mcp with Authorization: Bearer <contents of mcp-token>.
The terminal gesture cheat sheet
Every pane behaves the same, whichever CLI is inside it:
| Gesture | What it does |
|---|---|
| Click an unfocused pane | Focus it — start typing immediately |
| Scroll wheel | Scroll that pane's own transcript / scrollback |
| Option-drag | Select text (auto-copies to clipboard) |
| Right-click | Copy / Paste menu |
| ⌘C / ⌘V | Standard copy / paste |
The whole point: you never have to remember a different command per agent.
The Personal Agent slot
Point the "Personal Agent" setting at your own assistant CLI (command + args + a display name). It becomes a first-class orchestrator you can summon, and it receives an end-of-day handoff — a single prompt telling it to pull the day's activity and tidy the knowledge base. It's the seam where your own agent plugs into the cockpit; the app ships with the slot empty for you to fill.
Architecture (one paragraph)
src/main/ is the Electron main process (all the real logic — sessions, tmux, git, the MCP server,
the stores). src/preload/ is the typed IPC bridge. src/renderer/ is the React UI. src/shared/ is
pure, Node-free domain logic that carries the unit tests. Every subsystem has a shared/ half so its
logic is testable without touching live services. Start at src/main/index.ts.
License
MIT — see LICENSE. Use it, fork it, build your own cockpit.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。