Summary

Dulus provides a free, terminal-based AI agent, democratizing advanced LLM capabilities without API keys or subscriptions. It targets developers and power users by harvesting browser sessions or using local/paid providers, significantly reducing AI operational costs and complexity.

How It Works

This Python CLI agent uniquely captures guest browser sessions (Gemini, Claude.ai) as an LLM interface, also supporting 100+ paid providers via LiteLLM and local models via Ollama. Its core design emphasizes extensibility, integrating file manipulation, Bash, web browsing, and a novel plugin system that dynamically adapts any Python repository as a tool.

Quick Start & Requirements

Installation uses pip install dulus or OS-specific installer scripts (curl/powershell) with profile options ('full', 'standard', 'basic'), aiming for a 30-second setup. Docker is supported. Prerequisites include Python; optional dependencies cover GPU for local models, sounddevice/PortAudio for voice, and tkinter for Linux/WSL GUI. Links to the Dulus website and changelogs (/news command) are available.

Highlighted Details

• Zero-Cost Frontier AI: Leverages guest browser sessions and NVIDIA NIM's free tier (14 models, 40 RPM) to bypass API key requirements and costs.
• Universal Tool Integration: Features an "Auto-Adapter" plugin system dynamically integrating any Python repository as a tool without manifests, plus MCP support.
• Offline & Local First: Supports offline voice transcription (Whisper) and running local models (Ollama, LM Studio).
• Advanced Agent Capabilities: Includes multi-model debate, persistent semantic memory (Obsidian integration), sub-agents for parallel tasks, context compression, and session checkpointing.

Maintenance & Community

Maintained by single contributor KevRojo (@KevRojox). A community-launched Solana token ($DULUS) exists, with active development evident (v0.2.93, May 2026).

Licensing & Compatibility

Licensed under GPLv3, requiring derivative works to also be GPLv3, imposing copyleft restrictions on modifications and redistribution, especially concerning the --accept-all mode.

Limitations & Caveats

The Dulus OS sandbox is experimental. Voice features and Linux/WSL GUI require additional system library installations. The --accept-all permission mode bypasses prompts, posing a security risk. The project has a single primary maintainer, indicating a high bus factor.