Dulus — The Complete AI Harness
Inspiration
Dulus was born from a simple frustration: every AI tool was either too limited (closed SaaS) or too hard to use (raw APIs, complex SDKs). We wanted something that just works — install via pip, run a command, and have a fully capable AI agent with access to 100+ tools, 2000+ MCP servers, and 60,000+ skills. Named after the Cigua Palmera (our national bird 🇩🇴), Dulus represents freedom, adaptability, and reaching higher.
What it does
Dulus is a Python-native AI harness that transforms any LLM into a super-powered agent:
- Auto-Adapter: Drop any Python repo into Dulus and it becomes a native tool automatically
- 104+ built-in tools: Bash, file I/O, web search, browser automation, git, tmux, image OCR, and more
- 2000+ MCP servers: Via the Auto-Adapter — any MCP-compatible server plugs in instantly
- 60,000+ skills: From Anthropic marketplace, awesome-skills catalog, and Composio (1000+ apps)
- Multi-agent swarms: Spawn sub-agents in isolated git worktrees for parallel coding tasks
- Desktop GUI: Cross-platform app with embedded terminal, file manager, browser, and chat
- OpenAI-compatible endpoint: Self-host with 12 frontier models, your own API keys, your rules
- Fleet Platform: Control Room to orchestrate agent swarms across machines
- $Dulus token: First utility coin usable as AI compute/quota (launching soon)
How we built it
Pure Python, zero dependencies for core. The architecture is modular:
- Plugin system: Auto-discovers tools from any Python package via entry points
- Skill system: Reusable prompt templates with argument substitution
- MCP bridge: Native integration with the Model Context Protocol standard
- WebMCP support: Apps can expose tools that agents use directly on the page
- Nuitka builds: Compiled binaries for Windows/macOS/Linux with auto-updater
- Fleet server: FastAPI backend with SQLite, deployable anywhere
Built in the Dominican Republic 🇩🇴 by a single developer with an army of AI allies.
Challenges we ran into
- Binary size vs. features: Nuitka compiles are heavy — we split into core + plugin ecosystem
- Cross-platform paths: Windows POSIX shell quirks, OneDrive redirections, path escaping
- Real-time sync: Making the desktop GUI feel native while running Python in background
- API costs: Built local-first with $0 API fallbacks (OCR, search, etc.)
- Trust: Gained by shipping for 3+ months, 90K+ PyPI downloads, real customer 911s
Accomplishments that we're proud of
- 90,000+ PyPI downloads and counting
- 19,000+ MCP tools accessible via single command
- Auto-Adapter: Any Python repo → native Dulus tool, zero config
- Self-hosted endpoint: 12 frontier models, no vendor lock-in
- Real customers: Paying users, emergency fixes at 2AM, the whole journey
- Open source: GPL v3, community-driven, transparent
What we learned
- Composition > Construction: Being clever beats being funded
- Local-first wins: Users love $0 API options and data ownership
- Speed > Perfection: Ship fast, iterate with real users
- AI swarms work: Divide and conquer with specialized agents
What's next for Dulus
- $Dulus token launch: Utility coin for AI compute and platform quota
- Mobile app: iOS/Android harness
- Enterprise fleet: Multi-tenant agent orchestration
- More models: Integrate every frontier model as it drops
Try it out
pip install dulus
dulus --help
Or download the desktop app: github.com/Dulus-Ai/dulus-updates
Live demo: dulus.online
Built With
- dulus
- dulus-ai
- python
Log in or sign up for Devpost to join the conversation.