AgentForge

A multi-agent CLI that forges code through a Worker and an Evaluator collaborating in a loop — until the goal is achieved.

Inspired by OpenCode, AgentForge wraps AI coding CLIs (currently codex, with Claude and Gemini support planned) to orchestrate two specialized agents: one that acts, one that judges.

한국어 README

Demo

┌─ ⚙ WORKER AGENT ──────────────────────┬─ ◈ EVALUATOR AGENT ────────────────────┐
│                                        │                                         │
│ > Reading App.tsx...                   │ [Iter 1]                                │
│ > Writing dark_mode.css...             │ IMPROVE:                                │
│ > Modifying index.html...              │ Toggle button is missing.               │
│ ▌                                      │ Save state to localStorage.             │
│                                        │                                         │
│ [Iter 2]                               │ [Iter 2]                                │
│ > Adding ThemeToggle.tsx...            │ ✓ DONE                                  │
│ ✓ Created 2 files                      │ 결과물: ./src/ThemeToggle.tsx            │
│                                        │                                         │
└────────────────────────────────────────┴─────────────────────────────────────────┘
[AgentForge] > /plan Add dark mode to the React app

How It Works

User input (goal)
       │
       ▼
 ┌─────────────┐     code changes      ┌────────────────┐
 │   Worker    │ ───────────────────►  │   file system  │
 │   Agent     │   (full-auto sandbox) └────────────────┘
 └──────┬──────┘
        │ output
        ▼
 ┌──────────────────┐
 │   Evaluator      │
 │   Agent          │  (read-only sandbox — cannot modify files)
 └────────┬─────────┘
          │
          ├── DONE      →  Print Korean summary, wait for next command
          ├── IMPROVE   →  Send feedback to Worker → repeat
          └── REDIRECT  →  Change strategy entirely → repeat

The loop continues until the Evaluator decides DONE or the iteration limit is reached.

DONE Output Example

════════════════ ✓ Done — 3 iterations ════════════════

판단 이유
  Dark mode has been fully implemented with a toggle button.
  The theme state persists via localStorage across page reloads.

결과물 위치
  • ./src/ThemeToggle.tsx
  • ./src/App.tsx  (modified)

결과 요약
  React-based dark mode with persistent state. No extra dependencies.

Requirements

• Python 3.10+
• codex CLI — installed and authenticated
• Python packages: rich, prompt_toolkit

Installation

via npm (recommended)

npm install -g agentforge-multi

Dependencies (rich, prompt_toolkit) are installed automatically via postinstall.

via git

git clone https://github.com/<your-username>/AgentForge.git
cd AgentForge
bash install.sh

manually

cp agentforge ~/.local/bin/agentforge
chmod +x ~/.local/bin/agentforge
pip install rich prompt_toolkit

Usage

agentforge                  # Launch interactive CLI
agentforge -d /my/project   # Set working directory
agentforge -n 20            # Set max iterations (default: 5000)

Slash Commands

| Command | Description | |---------|-------------| | <goal text> | Send goal directly to the Worker agent and start the loop | | /plan <goal> | Plan Agent drafts a plan → Q&A → confirm → execute | | /exit | Exit AgentForge |

Type / to see available commands with autocomplete (like Claude Code).

/plan Flow

[AgentForge] > /plan Build a simple web page that says hello

[Plan Agent]
계획:
- Create index.html
- Add <h1>hello</h1>

accept (y/n) > y

▶ Starting Worker + Evaluator loop...

Options

| Flag | Default | Description | |------|---------|-------------| | -d DIR | . | Working directory | | -n N | 5000 | Max iterations | | --worker-model M | config default | Model for Worker agent | | --eval-model M | config default | Model for Evaluator agent |

Roadmap

• [x] codex CLI backend
• [ ] Claude CLI backend
• [ ] Gemini CLI backend
• [ ] Configurable agent personas
• [ ] Session history export

Project Structure

AgentForge/
├── agentforge      # Main executable script
├── install.sh      # Installation script
├── README.md       # English README (this file)
├── README.ko.md    # Korean README
└── .gitignore

Inspiration

• OpenCode — terminal-first AI coding agent
• Codex CLI — current underlying engine

License

MIT