Dulus - Open-source CLI AI Agent that can drive multi-model tool calling

What is Dulus

Dulus is an open-source CLI AI Agent of about 12K lines of Python, supporting 40+ models including Claude, GPT, Gemini, DeepSeek, Kimi, Qwen, and more. Tools can directly capture the guest session of browser-based AI web apps, turning them into tool-calling agents with 27 tools such as file reading and writing, code editing, Bash execution, and Web search. Dulus supports MCP, plugin hot-reloading, sub-agents, offline voice, persistent memory, and automatic snapshots, and provides PyQt6 GUI, WebChat, and Telegram multi-platform entry points.

Main features of Dulus

• Zero-key web session capture: automatically hijacks browser tabs such as Gemini, Claude.ai, Kimi.com, DeepSeek, and more, converting web conversations into a local Agent capable of tool calling.
• 27 built-in tools: covering Read, Write, Edit, Bash, Glob, Grep, WebFetch, WebSearch, NotebookEdit, GetDiagnostics, Memory, Tasks, Agents, Skills, and more.
• Multi-provider model switching: supports Anthropic, OpenAI, Google, DeepSeek, Kimi, Qwen, Zhipu, MiniMax, Ollama, LM Studio, and custom endpoints, with /model real-time switching during a session.
• MCP and plugin ecosystem: place .mcp.json to register any MCP server; Auto-Adapter can hot-load any Python repository as a plugin with zero configuration.
• Sub-agents and task management: generates sub-agents of types such as coder, reviewer, and researcher in independent git worktrees, collaborates through message passing, and includes a built-in task board.
• Offline voice and wake word: implements offline voice input based on Whisper-cpp, and voice output with Kokoro TTS; supports custom offline wake words such as "hey dulus".
• Persistent memory and snapshots: dual-scope memory (user-level + project-level) ranked by confidence and recency; automatically creates Checkpoints each round, supporting one-click rollback of conversations and file states.
• Multimodal interaction interfaces: provides four interaction methods: REPL terminal, Flask WebChat (accessible over LAN), PyQt6 native desktop GUI, and Telegram Bot.
• Developer enhanced mode: SSJ Mode provides debugging capabilities such as raw token counts, streaming latency timing, tool call inspector, and prompt injection viewer.

How to use Dulus

• Installation: run pip install dulus or curl -fsSL https://raw.githubusercontent.com/KevRojo/Dulus/main/install.sh | bash for one-click installation.
• Configure keys: set any provider environment variable such as export ANTHROPIC_API_KEY=sk-ant-..., or skip it entirely and use Ollama local models with browser hijacking mode.
• Start interaction: enter dulus in the terminal to enter the REPL, and use /model nvidia-web/deepseek-r1 or /model ollama/qwen2.5-coder to select a model.
• Enable zero-key mode: on first run, follow the wizard prompts to open the Gemini browser guest page; Dulus automatically captures the session, and tool calling can begin without an API Key.
• Execute tasks: directly enter natural language instructions such as "refactor the auth module", and the Agent automatically reads files, edits code, runs tests, and submits results.
• Extend capabilities: use /plugin install 仓库名@URL to hot-load any Python tool library, or /mcp Register external MCP server extensions.

Dulus's Core Advantages

• Truly Zero-Cost Startup: No API Key, credit card, or login required; get a full tool-calling Agent within 30 seconds through browser guest mode, lowering the barrier to entry to the extreme.
• Model Neutrality and Real-Time Switching: Not bound to a single provider, within the same session /model you can seamlessly switch between models such as Claude, DeepSeek, Kimi, and supports automatic fallback chains.
• Hot-Pluggable Ecosystem Expansion: Auto-Adapter turns any Python repository into a plugin in seconds; MCP and Composio provide 800+ skills, and ecosystem expansion requires no waiting for official adaptation.
• Full-Chain Offline Capabilities: Supports voice input and output, wake words, OCR, and Ollama local models, all of which can run in a completely disconnected environment, suitable for confidential networks and offline scenarios.
• Unified Core Across Multiple Interfaces: Terminal, Web, desktop GUI, and Telegram share the same Agent core and memory state, allowing users to freely switch interaction surfaces according to the scenario.

Dulus's Project Address

• Project Website：https://dulus.ai/
• GitHub Repository：https://github.com/KevRojo/Dulus

Comparison of Dulus with Similar Competitors

Dulus's Application Scenarios

• Personal Developer Coding Assistant: Replaces traditional IDE plugins, directly drives code reading, writing, refactoring, testing, and committing via natural language, and supports multi-model collaborative review.
• Zero-Budget AI Experiments: Students or beginners can use Gemini guest mode to freely experience the advanced tool-calling capabilities of cutting-edge models without subscribing to paid APIs.
• Offline/Confidential Environment Automation: Achieving fully privatized Agent workflows based on local Ollama models and offline speech in government, military, or corporate intranets.
• Telegram Remote Operations: Sending commands to Dulus at home or on a server via Telegram on a mobile phone to perform remote operations such as file management and log troubleshooting.
• Multi-Agent Collaborative Development: Assigning code writing, review, and testing to different sub-agents to advance in parallel, suitable for distributed development of large modules.