Multi-LLM orchestration with agentic execution, memory, and training data generation.

Puzld.ai is a terminal-native framework designed to orchestrate multiple Large Language Models (LLMs) for complex AI workflows. It empowers users to route tasks to specialized agents, compare their outputs, chain them in pipelines, and enable LLMs to safely explore and modify codebases. The framework aims to enhance productivity by automating multi-step AI processes and generating valuable training data.

How It Works

The core of Puzld.ai is its agent orchestration engine, which intelligently routes tasks to the most suitable LLM agent based on capabilities. It supports advanced interaction patterns including parallel comparison, sequential pipelines, and multi-agent collaboration (correction, debate, consensus). A key differentiator is its "Agentic Mode," which grants LLMs controlled access to tools for exploring codebases (view, glob, grep, bash) and proposing file edits, all subject to explicit user permission prompts. This approach allows LLMs to act as sophisticated coding assistants while maintaining user control over modifications.

Quick Start & Requirements

• Installation: npm install -g puzldai or npx puzldai.
• Prerequisites: Node.js (npm/bun), and the official CLIs for desired agents (e.g., Claude CLI, Gemini CLI, Ollama). Some agent CLIs may have specific requirements (e.g., GPU for local models).
• Links: Project repository: https://github.com/MedChaouch/Puzld.ai.git.

Highlighted Details

• Agentic Mode: LLMs can explore codebases using tools like glob, grep, and bash, and propose file edits (write, edit) with granular, permission-based prompts for user approval.
• Multi-Agent Collaboration: Features include correct (producer-reviewer), debate (argumentation), and consensus (voting) modes to refine outputs.
• Observation Layer: Logs all agentic interactions, including inputs, outputs, and decisions, to facilitate the generation of DPO (Direct Preference Optimization) training data.
• Codebase Indexing: Utilizes Abstract Syntax Tree (AST) parsing and semantic search to enable LLMs to understand project structure and context.

Maintenance & Community

The project is developed by Med Chaouch. No specific community channels (e.g., Discord, Slack) or detailed roadmap information are provided in the README.

Licensing & Compatibility

• License: AGPL-3.0-only.
• Compatibility: This strong copyleft license requires any derivative works to be distributed under the same AGPL-3.0-only license. This may impose significant restrictions on integration with proprietary or closed-source software.

Limitations & Caveats

• Certain agent CLIs (Gemini, Codex) may bypass PuzldAI's permission system for file reading due to their internal implementations, unlike Claude and Ollama which fully respect the prompt system.
• The AGPL-3.0-only license's copyleft nature necessitates careful consideration for commercial adoption or integration into closed-source projects.