A developer has introduced Dulus — a lightweight Python implementation of Claude Code that is not tied to a single provider. The tool supports over 100 backends via LiteLLM, local models through Ollama, and includes over 2,000 MCP servers, providing universal AI access via CLI, Web, GUI, and Telegram.
What Happened
Dulus, an open-source framework acting as a multimodal AI agent, has been introduced. The project supports integration with numerous providers via LiteLLM, works with local models through Ollama, and features built-in capabilities such as Whisper voice input, OCR, and multi-agent systems (sub-agents). The tool operates through various interfaces: terminal (CLI), web interface, graphical user interface (GUI), and Telegram.
Context
Dulus solves the problem of vendor lock-in, allowing users to switch between commercial APIs (such as Claude or GPT) and local solutions without rewriting code. The system's scalability is ensured by support for the Model Context Protocol (MCP), turning the tool into a full-fledged operating environment for agents with access to a massive ecosystem of tools.
Why It Matters for the Industry
The project demonstrates a key trend toward the decentralization and multimodality of AI agents. It provides developers with a ready-made architectural skeleton for creating complex agentic systems with voice and vision support, lowering the barrier to entry and reducing dependence on the closed ecosystems of large vendors.
Why It Matters for Users
Users can use Dulus as a powerful personal coding agent integrated into their familiar workflows via CLI or Telegram. Thanks to support for local models through Ollama or NVIDIA NIM, advanced features can be used without the need for constant cloud API key payments, while maintaining a high level of privacy.
What Is Not Yet Known / Limitations
At the current stage, the project is being developed by a single developer, so official benchmarks are unavailable. Corporate users and ML engineers should exercise caution and conduct security audits, especially regarding code execution in sandboxes.
Sources
Author
Look at AI, Editorial Team