openmax 0.2.1

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

openMax

Multi AI Agent orchestration hub

---

One command, multiple AI agents, one window.

openMax breaks down your task, dispatches agents (Claude Code, Codex, OpenCode) into Kaku terminal panes, monitors progress, and intervenes when needed.

Architecture

openmax run "Build a blog with Next.js"
                    |
                    v
   +--------------------------------+
   |          Lead Agent            |
   |  (powered by claude-agent-sdk) |
   |                                |
   |  Align -> Plan -> Dispatch ->  |
   |  Monitor -> Review -> Report   |
   +---------------+----------------+
                   |
                   v
   +---------------+----------------+
   |        Kaku Terminal Window    |
   |  +-------------+------------+ |
   |  | claude-code | codex      | |
   |  | components  | API routes | |
   |  +-------------+------------+ |
   |  | claude-code | opencode   | |
   |  | tests       | styling    | |
   |  +-------------+------------+ |
   +--------------------------------+

Install

pip install openmax

Requirements:

• macOS (Kaku is macOS-only)
• Python 3.10+
• Kaku terminal — auto-prompted via brew install --cask kaku if missing
• At least one agent CLI: Claude Code (claude), Codex (codex), or OpenCode (opencode)

Quick Start

# Basic usage — give it a task and let it work
openmax run "Build a blog with Next.js"

# Specify a working directory
openmax run "Add authentication to the app" --cwd ~/projects/my-app

# Use a specific model for the lead agent
openmax run "Refactor the database layer" --model claude-sonnet-4-20250514

# Keep panes open after completion for manual inspection
openmax run "Fix all failing tests" --keep-panes

# Only use Claude Code and Codex (first one is preferred)
openmax run "Build the API" --agents claude-code,codex

# Force all tasks to use a single agent type
openmax run "Fix lint errors" --agents claude-code

Usage

openmax run

The core command. The lead agent (powered by claude-agent-sdk) will:

1. Align — clarify your goal, identify scope
2. Plan — decompose into parallelizable sub-tasks
3. Dispatch — spawn agents into Kaku panes (one window, auto grid layout)
4. Monitor — read agent output, intervene if stuck or off-track
5. Review — cross-check deliverables, run tests, verify integration
6. Report — summarize results, ensure changes are committed

openmax run "Build a REST API with FastAPI and SQLAlchemy"

All options:

Option	Description	Default
--cwd PATH	Working directory for agents	Current directory
--model MODEL	Model for the lead agent	claude-sonnet-4-20250514
--max-turns N	Max lead agent loop turns	50
--keep-panes	Don't close panes on exit	false
--session-id ID	Persistent session identifier	Auto-generated
--resume	Resume a persistent session (requires --session-id)	false
--agents LIST	Comma-separated allowed agent types, in preference order	All available

openmax doctor

Check that your environment is ready before running tasks.

$ openmax doctor
openMax environment check
──────────────────────────────────────────
  ✅  Python              3.11.4
  ✅  Kaku CLI            0.6.0
  ✅  claude              2.1.76
  ❌  opencode
       Fix: See https://github.com/opencode-ai/opencode  (optional)
──────────────────────────────────────────
1 issue found.

openmax validate-config

Validate your agent TOML configuration:

$ openmax validate-config --cwd ~/projects/my-app
Validating agent config for /Users/me/my-app
  Built-in: claude-code, codex, opencode, generic  ✅
  .openmax/agents.toml
    remote-codex  ✅  ssh devbox bash -lc ...
    broken-agent  ❌  command must be a non-empty string array

openmax panes

List all Kaku terminal panes and their status.

$ openmax panes
Found 4 panes:
  Pane 1: claude-code (120x30) ★
  Pane 2: codex (120x30)
  Pane 3: claude-code (120x30)
  Pane 4: opencode (120x30)

openmax read-pane

Read the terminal output of a specific pane — useful for checking what an agent is doing.

$ openmax read-pane 1

openmax memories

Show learned workspace memory from previous runs.

# Show memory for the current directory
$ openmax memories

# Show memory for a specific workspace
$ openmax memories --cwd ~/projects/my-app

# Limit number of entries
$ openmax memories --limit 5

openmax recommend-agents

Show ranked agent recommendations for a task based on past runs in the workspace.

$ openmax recommend-agents "Build REST API" --cwd ~/projects/my-app
Agent recommendations for Build REST API
- codex: 12
  Converged fastest on API work in last 3 runs
- claude-code: 8
  Good for general coding tasks

Session Resume

Resume a previous session to continue where you left off:

# Start a named session
openmax run "Build the frontend" --session-id my-frontend

# Later, resume it
openmax run "Continue the work" --session-id my-frontend --resume

Examples

# Full-stack app development
openmax run "Build a todo app with React frontend and Express backend"

# Bug fixing across a codebase
openmax run "Find and fix all TypeScript type errors in src/"

# Code refactoring
openmax run "Migrate all class components to functional components with hooks"

# Testing
openmax run "Write comprehensive unit tests for the utils/ directory"

# Multi-language projects
openmax run "Add Python bindings for the Rust core library"

Supported Agents

Agent	Command	Notes
Claude Code	claude	Best for most coding tasks
Codex	codex	OpenAI Codex CLI
OpenCode	opencode	OpenCode CLI

All agents run interactively in their own pane by default. You can click into any pane and type to intervene at any time. The lead agent also monitors and sends corrections automatically.

You can also define your own agents in config, including arbitrary local CLIs or remote entrypoints over SSH.

Agent Selection

By default, the lead agent automatically picks the best agent for each sub-task (claude-code is the default). Use --agents to restrict and prioritize built-in or configured agent names:

# Prefer Codex, fall back to Claude Code if needed
openmax run "Build the API" --agents codex,claude-code

# Only use Claude Code for everything
openmax run "Refactor auth module" --agents claude-code

The order matters — the first agent in the list is the preferred default. If the lead agent tries to use an agent not in the list, it automatically falls back to the first one.

The lead agent also learns from past runs. After completing a task, it stores what worked well via remember_learning. Future runs in the same workspace will receive these recommendations automatically. View them with openmax memories.

Custom Agent Config

openMax merges agents from these locations, in this order:

1. Built-in agents
2. ~/.config/openmax/agents.toml
3. <workspace>/.openmax/agents.toml
4. OPENMAX_AGENTS_FILE=/path/to/agents.toml

List the effective agent registry for a workspace with:

openmax list-agents --cwd /path/to/repo

Example config:

[agents.remote-codex]
command = ["ssh", "devbox", "bash", "-lc", "cd {cwd_sh} && codex"]
interactive = true
startup_delay = 8

[agents.remote-review]
command = ["ssh", "devbox", "bash", "-lc", "codex exec {prompt_sh}"]
interactive = false

Supported placeholders:

• {cwd} / {prompt}: raw substitution
• {cwd_sh} / {prompt_sh}: shell-escaped substitution for commands such as ssh ... "cd ... && tool"

Notes:

• Interactive agents start the command first, then openMax sends the task prompt into the pane.
• Non-interactive agents must include {prompt} or {prompt_sh} in command.
• Use startup_delay for slow-starting commands such as SSH sessions or remote shells.

How It Works

openMax uses a lead agent that has no direct access to files or code. Instead, it orchestrates through MCP tools:

Tool	Purpose
dispatch_agent	Spawn an agent in a Kaku pane
read_pane_output	Check what an agent is doing
send_text_to_pane	Send follow-up instructions
list_managed_panes	Get pane states
mark_task_done	Track sub-task progress
report_completion	Finalize and report results

Cleanup: On exit (normal completion, Ctrl-C, or SIGTERM), all managed panes are killed automatically. Use --keep-panes to keep them open.

Best Practice: Let Your Agent Call openMax

The most powerful way to use openMax is to let your AI agent delegate work to it asynchronously. Instead of running openMax manually, instruct your agent (Claude Code, Cursor, etc.) to spawn openMax as a background process:

# In your agent's prompt or CLAUDE.md:
"When you need to parallelize work across multiple files or modules,
 run openmax in the background and continue with other tasks."

Example workflow — your agent is building a full-stack app:

# Your agent runs this in the background, then continues its own work
openmax run "Build React components for dashboard" --cwd ./frontend --agents claude-code &

# Meanwhile, your agent works on the backend itself
# ...

# Later, check the results
openmax read-pane 1

This turns a single agent into a team of agents — your primary agent handles the high-level architecture while openMax manages the parallel sub-tasks. The key benefits:

• Async by nature — openMax runs independently; your agent doesn't block
• Automatic monitoring — the lead agent watches all sub-agents, intervenes when stuck
• Clean separation — each agent works in its own terminal pane, no conflicts

License

MIT