KevRojo/Dulus

What it solves

Dulus is a provider-independent agent runtime that removes the barrier to entry for frontier AI. It allows users to run powerful AI agents without needing API keys, credit cards, or corporate accounts by leveraging web-session harvesting to talk to models like Gemini, Claude, and ChatGPT as if the user were interacting with their web interface.

How it works

• Web-Session Engine: Uses headless browsers to capture anonymous or authenticated web sessions, allowing the agent to communicate with frontier models via their own web protocols instead of paid APIs.
• Provider Router: A unified interface that supports over 34 providers, including direct cloud APIs, local models (Ollama, LM Studio), and browser-backed sessions.
• Python Console: A persistent REPL (heap) that stores large datasets in a kernel's memory across turns, preventing the need to re-send massive amounts of data in the conversation context window.
• Lookback: A sliding window mechanism that sends only recent turns to the API while archiving the full history locally, significantly reducing token costs and maintaining prompt cache hits.
• Auto-Adapter: Dynamically inspects Python repositories to infer operations and generate plugin tools on the fly.

Who it’s for

Developers and AI enthusiasts who want a free, local-first, or provider-agnostic way to run agentic AI with extensive tool integration (MCP, skills, plugins) across various interfaces (CLI, desktop app, Telegram).

Highlights

• Zero-Key Access: Access to frontier models via web sessions without API keys.
• Multimodal Tooling: Native support for MCP servers, a custom Auto-Adapter for Python repos, and a library of skills.
• Token Efficiency: Features like the Python console and Lookback minimize token usage and cost.
• Advanced Agent Orchestration: Supports sub-agents in isolated worktrees and "Mesa Redonda" for multi-model debates.
• Flexible Deployment: Runs via pip, Docker, or as a standalone desktop application.