openmax 0.1.3

Multi AI Agent orchestration hub — dispatch interactive AI agents across terminal panes

openMax

Multi AI Agent orchestration hub — dispatch interactive AI agents across terminal panes.

openMax is a lead-agent system that decomposes your task, spawns multiple AI coding agents (Claude Code, Codex, OpenCode, etc.) in separate Kaku terminal windows, monitors their progress, and reports results — all automatically.

How it works

openmax run "用 Next.js 写个博客"
         │
         ▼
┌──────────────────────────────────────────────┐
│  Lead Agent (claude-agent-sdk)               │
│  Phase 1: Align goal                         │
│  Phase 2: Plan & decompose                   │
│  Phase 3: Dispatch agents                    │
│  Phase 4: Monitor & correct                  │
│  Phase 5: Summarize & report                 │
└──────────┬───────────────────────────────────┘
           │ kaku spawn --new-window
           ├──────────────┬──────────────┐
           ▼              ▼              ▼
   ┌──────────────┐ ┌──────────┐ ┌──────────┐
   │ Window 1      │ │ Window 2  │ │ Window 3  │
   │ claude-code   │ │ codex     │ │ claude    │
   │ "写组件"      │ │ "写API"   │ │ "写测试"  │
   └──────────────┘ └──────────┘ └──────────┘
         ↑              ↑              ↑
     interactive    interactive    interactive
     (用户可直接点进任意窗口干预 agent)

Install

pip install openmax

or

uv pip install openmax

Prerequisites

• Python 3.10+
• Kaku terminal (based on WezTerm)
• At least one AI agent CLI installed:
  • Claude Code (claude)
  • Codex (codex)
  • OpenCode (opencode)

Usage

Run a task

openmax run "用 Next.js 写一个博客系统"

The lead agent will:

1. Restate and clarify your goal
2. Decompose it into parallelizable sub-tasks
3. Open a new Kaku window for each sub-task with the appropriate agent
4. Monitor each agent's output and intervene if needed
5. Report completion with cost/token summary

Options

# Specify working directory
openmax run "重构 API 模块" --cwd /path/to/project

# Use a specific model for the lead agent
openmax run "写单元测试" --model claude-sonnet-4-20250514

# Limit agent loop turns
openmax run "修复 bug" --max-turns 30

List panes

openmax panes

Read a pane's output

openmax read-pane <pane_id>

Supported agents

Agent	Type	Mode	Command
Claude Code	claude-code	Interactive	claude
Codex	codex	Interactive	codex
OpenCode	opencode	Interactive	opencode
Generic	generic	Interactive	claude

All agents run interactively — you can click into any Kaku window and type to intervene directly.

Architecture

src/openmax/
├── cli.py              # CLI entry point (click)
├── lead_agent.py       # Lead agent: task decomposition + monitoring
│                       #   Uses claude-agent-sdk with custom MCP tools
├── pane_manager.py     # Kaku window/pane lifecycle management
├── kaku.py             # Kaku CLI availability check
└── adapters/
    ├── base.py             # AgentAdapter ABC + AgentCommand
    ├── claude_code.py      # Claude Code adapter
    ├── codex_adapter.py    # Codex adapter
    ├── opencode_adapter.py # OpenCode adapter
    └── subprocess_adapter.py # Generic CLI adapter

Lead agent tools

The lead agent has 6 custom tools (via in-process SDK MCP server):

Tool	Description
dispatch_agent	Open a new Kaku window and start an agent with a prompt
read_pane_output	Read an agent's terminal output to check progress
send_text_to_pane	Send text to an agent (follow-up instructions, intervention)
list_managed_panes	List all managed panes and their states
mark_task_done	Mark a sub-task as completed
report_completion	Report overall completion percentage

Lifecycle

• On completion or Ctrl-C, all managed Kaku windows/panes are automatically closed
• SIGINT and SIGTERM are handled for clean shutdown
• The PaneManager tracks window IDs, pane states (idle/running/done/error), and handles cleanup

License

MIT