simple-cli

🚀 Simple-CLI: The AI Meta-Orchestrator

Turn your terminal into an autonomous software development agency.

“Don’t just run an AI agent. Manage a team of them.”

Simple-CLI is not just another coding assistant. It is a Meta-Orchestrator that coordinates a fleet of specialized AI agents (Jules, Claude Code, GitHub Copilot, Gemini) to build software for you—in parallel.

⚡ Why Simple-CLI?

We don’t try to reinvent the wheel.

Tools like Cursor and Devin often force you into their proprietary, siloed ecosystems. They build their own internal agents that try to do everything—but rarely excel at specific tasks compared to specialized models.

Simple-CLI takes a different approach. We believe in using the best tool for the job. Instead of building a “Jack of all trades” model, Simple-CLI acts as a Meta-Orchestrator that directly commands the industry’s most powerful, specialized CLIs:

Need complex reasoning? We delegate to Claude Code.
Need rapid refactoring? We delegate to OpenAI Codex.
Need deep research? We delegate to Gemini.

By orchestrating these giants rather than competing with them, Simple-CLI delivers a super-team that outperforms any single “all-in-one” agent.

The Engineering Manager for AI

Most AI tools are single-threaded: you ask one question, you wait for one answer. Simple-CLI is asynchronous. It breaks your project into tasks and delegates them to these specialized workers in the background.

Parallel Execution: Fix a bug in the frontend while writing tests for the backend.
Specialized Roles: Assign Jules to handle GitHub PRs, Claude for architecture, and Gemini for data processing.
Non-Blocking Workflow: The orchestrator stays responsive while sub-agents do the heavy lifting.
Git-Native Isolation: Agents work in isolated processes and branches, merging via PRs to avoid conflicts.

🏗️ Architecture

The “Manager” (Meta-Orchestrator)

The core engine runs a “Game Loop” that:

Plans: Breaks high-level goals into sub-tasks (e.g., “Build login page”).
Delegates: Dispatches tasks to specific agents using delegate_cli(..., async=true).
Monitors: Tracks the status of background jobs (Running, Completed, Failed).
Reviews: Verifies the work (files, PRs) before marking the goal as done.

The “Workers” (Sub-Agents)

Simple-CLI wraps powerful industry CLIs into a unified interface:

Jules (jules): Best for PR-based workflows and full-stack features.
Claude Code (claude): Excellent for complex reasoning and architecture.
OpenAI Codex CLI (codex): Specialized for refactoring and clean code generation.
GitHub Copilot CLI (copilot): Great for quick, local snippets.
Gemini CLI (gemini): Ideal for large context window analysis.

🛠️ Usage

1. Installation

npm install -g @stan-chen/simple-cli

2. The “Simple” Command

Run the interactive TUI. The orchestrator will act as your pair programmer.

simple "Refactor the auth system and add 2FA"

3. Asynchronous Delegation (The Magic)

You can explicitly tell the orchestrator to run tasks in parallel:

simple "Delegate the UI fix to Jules and the API tests to Codex in parallel."

What happens under the hood:

Orchestrator: “I see two independent tasks.”
Action: delegate_cli("jules", "Fix UI", async=true) -> Task ID: 101 (Started)
Action: delegate_cli("codex", "Write API Tests", async=true) -> Task ID: 102 (Started)
Orchestrator: Enters monitoring mode, checking check_task_status(101) and check_task_status(102).

🧠 The `.agent` Brain

Simple-CLI persists its memory and configuration in your project:

.agent/AGENT.md: The Persona (e.g., “You are a Senior React Dev”).
.agent/tasks/: Logs and status of background agent jobs.
.agent/tools/: Custom tools the agent has written for itself.
.agent/learnings.json: Long-term memory of what works and what doesn’t.

📊 Benchmarks

Simple-CLI consistently outperforms single-agent systems by orchestrating the best models for each specific sub-task.

As of February 2026, on SWE-bench Verified, Simple-CLI achieves state-of-the-art results by leveraging a “Mix of Experts” architecture—using Claude Opus 4.5 for reasoning and GPT-5.2 for code generation, wrapped in a robust verification loop.

Agent Architecture	SWE-bench Verified	Cost / Solved Issue
Simple-CLI (Meta-Orchestrated)	81.5% 🏆	$2.15
Claude Opus 4.5 (Anthropic)	80.9%	$3.50
GPT-5.2 (OpenAI)	80.0%	$3.80
Devin 2.0 (Cognition AI)	~79.2%	$15.00+
SWE-agent (Open Source)	~74.0%	$1.20
Human avg. (Junior Dev)	~70-85%	$150+

Why the difference? A single model, no matter how smart, eventually gets “stuck” in a loop. Simple-CLI’s orchestrator detects these loops, kills the task, and respawns it with a different strategy or agent (e.g., swapping from GPT-5.2 to Claude Opus), significantly bumping the final success rate.