Meld

Multi-model planning convergence CLI that synthesizes plans from multiple AI providers.

Overview

Meld orchestrates multiple AI advisors (Claude, Gemini, OpenAI) to collaboratively develop implementation plans. Through iterative refinement, the advisors converge on a unified approach that incorporates diverse perspectives.

Installation

pip install meld

Or install from source:

git clone https://github.com/meld-project/meld.git
cd meld
pip install -e .

Usage

Basic Usage

# Run with a task description
meld "Build an authentication system with OAuth2 support"

# Read task from file
meld --file task.txt

# Pipe task from stdin
echo "Design a REST API for user management" | meld

With PRD Context

meld "Implement user authentication" --prd requirements.md

Options

meld "Task description" \
  --rounds 5              # Max iteration rounds (default: 5)
  --timeout 600           # Timeout per advisor in seconds (default: 600)
  --output plan.md        # Write final plan to file
  --json-output summary.json  # Write JSON summary
  --quiet                 # Minimal output, no TUI
  --verbose               # Include raw advisor outputs
  --skip-preflight        # Skip environment checks

Session Management

# Resume an interrupted run
meld --resume 2026-01-16T02-47-17Z-abc123

# Custom run directory
meld "Task" --run-dir ./custom-runs/

Environment Check

# Check if all required CLIs are available
meld doctor

How It Works

1. Preflight: Verifies required CLI tools (claude, gemini, openai) are available
2. Initial Planning: Each advisor independently generates an implementation plan
3. Critique Rounds: Advisors review and critique each other's plans
4. Synthesis: A "melder" synthesizes critiques into a unified plan
5. Convergence: Process repeats until advisors reach consensus or max rounds

Requirements

• Python 3.10+
• At least one AI CLI tool:
  • claude (Anthropic Claude)
  • gemini (Google Gemini)
  • openai (OpenAI)

License

MIT