multiagent-terminal

Run multiple Claude Code agents in parallel. Each agent gets its own git worktree, terminal, and status tracking.

Why

Claude Code in one terminal flickers, can't scroll while it's thinking, gives no signal when it's done, and forgets what you were doing. Run ten of them at once, each in its own git worktree, each with a clean terminal that scrolls and notifies — and stop losing context to the boring stuff.

Features

• No flickering --- unlike raw Claude Code terminal, the UI never flickers or redraws
• Scroll-safe --- reading history while agent outputs? Viewport stays put, no scroll hijacking
• Session persistence --- agents resume where they left off after app restart via --continue
• Parallel agents --- spawn unlimited Claude sessions, each in its own git worktree
• Git worktrees --- full isolation between agents, no branch conflicts, clean parallel work
• Live status --- thinking/working/done detection via PTY output analysis and named pipes
• Notifications --- system alerts + green/yellow card highlights when agents finish or are working
• Terminal per agent --- full xterm.js with GPU acceleration, scrollback, and bracketed paste
• Git-aware --- branch display, file count, line stats (+/-), PR detection, diff viewer
• Shell tab --- auxiliary terminal per agent for manual git/shell commands
• Git log --- commit history with stats, click to view individual diffs
• Cross-agent search --- Ctrl+F across every agent's terminal log, trigram-indexed in a worker thread
• Token spend dashboard --- per-agent and aggregate cost/usage with daily charts
• Cross-platform --- Linux, macOS, Windows (named pipes adapt per OS)
• Scope --- only Claude Code and GitHub are supported. Other AI agents (Cursor, Gemini CLI, etc.) and other Git hosts (GitLab, Bitbucket) are not — status tracking parses Claude's statusline, PR detection goes through gh.

Install

Prebuilt (Linux)

Download from the latest release:

• AppImage — chmod +x multiagent-terminal-*.AppImage && ./multiagent-terminal-*.AppImage
• Debian/Ubuntu — sudo dpkg -i multiagent-terminal_*_amd64.deb

From source

git clone https://github.com/ClickHouse/multiagent-terminal.git
cd multiagent-terminal
npm install
npm run dev

Requirements

The app checks these on startup and won't launch without them:

Dev

npm run dev      # electron-vite dev server with HMR
npm run build    # production build to out/
npx electron . --no-sandbox   # run the built bundle without packaging
npm run dist:linux            # produce AppImage + .deb in release/
npm run dist:mac              # macOS dmg
npm run dist:win              # Windows nsis installer

How it works

Each agent is isolated

your-repo/                    # base repository
your-repo-agent-1/            # worktree for agent 1
your-repo-agent-2/            # worktree for agent 2
your-repo-fix-auth/           # worktree for agent 3

Agents don't interfere with each other. Each gets a branch, a working directory, and a Claude process.

Status detection

The status state machine tracks what each agent is doing without polling Claude:

idle ---(Enter)--> thinking ---(200B output)--> working ---(3s silence)--> idle
                       |                                                    ^
                       +-----------(60s timeout)----------------------------+

A named pipe (mkfifo on Unix, \\.\pipe\ on Windows) receives JSON status updates from Claude's statusline, providing model name, context usage, and cost.

What you see

Done cards stay green until you click them. The selected card flashes green briefly then fades back. Active cards glow amber. Everything is instant --- no transition animations.

Stack

Settings

License

Apache 2.0 — see LICENSE.md.