Multi-agent orchestration platform for Gemini CLI, Claude Code, Codex, and Qwen Code — 39 specialists, parallel subagents, persistent sessions, and built-in code review, debugging, security, SEO, accessibility, and compliance tools
Maestro is a multi-agent development orchestration platform with 39 specialists, an Express path for simple work, a 4-phase standard workflow for medium and complex work, persistent session state, and standalone review/debug/security/perf/seo/accessibility/compliance entrypoints. It runs from one canonical src/ tree across Gemini CLI, Claude Code, Codex, and Qwen Code.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
git clone https://github.com/josstei/maestro-orchestrate
cd maestro-orchestrate
qwen extensions link .
Verify with qwen extensions list. Qwen Code uses the same /maestro:* command surface as Gemini CLI and reads QWEN.md as its context file.
Quick Start
Start a full orchestration with the runtime-specific entrypoint:
| Runtime | Example |
|---------|---------|
| Gemini CLI | /maestro:orchestrate Build a REST API for a task management system with user authentication |
| Claude Code | /orchestrate Build a REST API for a task management system with user authentication |
| Codex | $maestro:orchestrate Build a REST API for a task management system with user authentication |
| Qwen Code | /maestro:orchestrate Build a REST API for a task management system with user authentication |
Maestro classifies the task, chooses Express or Standard workflow, asks the required design questions, produces an implementation plan when needed, delegates execution to specialists, runs a quality gate, and archives the session state in docs/maestro/.
Examples
Usage examples: EXAMPLES.md. Gemini/Qwen forms shown:
Full orchestration: /maestro:orchestrate Build a REST API for a task management system with user authentication
Standalone review: /maestro:review Review the staged changes for correctness, regressions, security, maintainability risk, and missing tests
Security audit: /maestro:security-audit Audit authentication, authorization, data exposure, secret handling, and exploitability risks
Configuration
Defaults work; these settings tune behavior:
| Setting | Default | Purpose |
|---------|---------|---------|
| MAESTRO_STATE_DIR | docs/maestro | Session, plan, and archive output path |
| MAESTRO_EXECUTION_MODE | ask | Choose parallel, sequential, or prompt |
| MAESTRO_AUTO_ARCHIVE | true | Archive successful sessions automatically |
| MAESTRO_MAX_RETRIES | 2 | Retry limit for failed phases |
| MAESTRO_MAX_CONCURRENT | 0 | Parallel-agent cap, where 0 means no Maestro cap |
| MAESTRO_DISABLED_AGENTS | unset | Specialists to exclude from assignment |
For Claude Code and Codex, Maestro intentionally avoids bare skill names that collide with host commands. Use /review-code, /debug-workflow, and /resume-session in Claude Code, and $maestro:review-code, $maestro:debug-workflow, and $maestro:resume-session in Codex, so built-in /review, /debug, and /resume commands keep working.
Qwen Code uses the same /maestro:* command surface as Gemini CLI.
Workflow
Express: For simple work. Maestro asks 1-2 clarifying questions, proposes a brief, delegates to one specialist, runs code review, and archives without a design doc or implementation plan.
Standard: For medium and complex work. Maestro runs Design, Plan, Execute, and Complete phases with explicit approval gates, phased execution, and final review blocking on unresolved Critical or Major findings.
Outputs and Success Criteria
Maestro writes orchestration outputs under MAESTRO_STATE_DIR, usually docs/maestro/. Standard workflow outputs include active session state, design documents, implementation plans, phase reports, validation output, and archived records.
A successful run must have an approved plan when Standard workflow is used, completed phase reports, validation results for the changed surface, and no unresolved Critical or Major review findings. If a phase cannot complete, Maestro records the blocker and the next required action instead of silently continuing.
Security and Permissions
Maestro follows the host runtime's tool permissions, sandboxing, and confirmation model. It does not require committed secrets or long-lived credentials, and orchestration session state stays inside MAESTRO_STATE_DIR unless configured otherwise. Use MAESTRO_DISABLED_AGENTS to restrict specialists in sensitive repositories, and run $maestro:security-audit or the equivalent runtime command before adopting changes that touch authentication, authorization, secrets, or data exposure paths.
Documentation
EXAMPLES.md for copyable usage scenarios across all runtimes
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
