Hermes

Run a fleet of AI coding agents in parallel — each in its own isolated git worktree, with real-time streaming, code review, and one-click merge.

Hermes is a self-hosted macOS desktop app that lets you orchestrate multiple AI agents (Claude Code, GPT, MiniMax) working simultaneously across your repos. Add a repository, spin up workspaces, assign tasks to different models, watch them work in real time, review the diffs, and merge to main.

Features

• Parallel agent execution — run multiple Claude Code, GPT, or MiniMax agents at the same time
• Git worktree isolation — each workspace gets its own branch, no conflicts between agents
• Multi-model support — Claude for coding, GPT for research, MiniMax for general tasks, all through a unified proxy
• Real-time streaming — watch agent output live as they work (SSE)
• Code review & merge — diff viewer with approve/reject/merge workflow, conflict detection
• Mission control dashboard — see all agents across all repos in one grid view
• Cost tracking — token usage and cost estimates per model, per run
• SQLite persistence — all state survives restarts, no external database needed
• Native macOS app — Electron with tray icon, keyboard shortcuts, hiddenInset title bar

Quick Start

cd App
npm install
cd frontend && npm install && npm run build && cd ..
npm start

Development Mode

cd App
npm install
cd frontend && npm install && cd ..
npm run dev

Architecture

App/
  main.js                    # Electron main process
  backend/
    server.js                # Express API server
    db.js                    # SQLite database
    proxy/router.js          # Unified model proxy with cost tracking
    agents/
      claude-agent.js        # Claude Code CLI spawner
      gpt-agent.js           # OpenAI API agent
      minimax-agent.js       # MiniMax API agent
    git/
      worktree.js            # Git worktree lifecycle
      review.js              # Diff, merge, conflict detection
    streaming/sse.js         # Server-Sent Events
  frontend/                  # React + Vite + Tailwind
    src/
      components/            # Sidebar, Chat, Terminal, DiffViewer
      stores/                # Zustand state management
      hooks/                 # SSE + API hooks

Environment Variables

ANTHROPIC_API_KEY=sk-...     # For Claude Code CLI
OPENAI_API_KEY=sk-...        # For GPT research agents
MINIMAX_API_KEY=...          # For MiniMax general agents
HERMES_PORT=8765             # Backend port (default: 8765)

API

Method	Endpoint	Description
GET	/api/repos	List repositories
POST	/api/repos	Add repository (clone from URL)
POST	/api/workspaces	Create workspace (git worktree)
POST	/api/workspaces/:id/spawn	Spawn agent { model, task }
POST	/api/workspaces/:id/kill	Stop running agent
GET	/api/workspaces/:id/diff	Get diff vs main branch
POST	/api/reviews/:id/merge	Merge agent's work to main
GET	/api/dashboard	All agents across all repos
GET	/api/proxy/usage	Cost & token tracking
GET	/events	SSE stream

Roadmap

• Multi-feature parallel build — Orchestrate N agents writing independent modules to the same repo, then auto-merge and deploy. Use case: "Build 10 new features across a codebase simultaneously, each agent in its own worktree, with conflict resolution and single-push deploy."
• Task templates — Pre-built task definitions (e.g., "Add feature with JS + CSS + HTML wiring") that agents can execute with minimal prompting
• Agent-to-agent handoff — Let one agent's output feed into another's input (e.g., research agent → implementation agent → test agent)
• Deployment hooks — Auto-deploy to Vercel/Netlify/etc. after successful merge
• Cost budgets — Set per-task or per-session cost limits, auto-pause agents approaching budget
• Diff review queue — Batch review mode for reviewing outputs from many agents at once
• CI integration — Run tests automatically on each agent's worktree before allowing merge

Inspiration

Hermes was inspired by Conductor — a fantastic app for running parallel Claude Code agents. Hermes takes the same core concept and extends it with multi-model support (Claude + GPT + MiniMax), a unified cost tracking proxy, and a self-hosted architecture with no cloud dependency.

License

MIT