mini-claw is a local-first, agentic AI development framework built around the Model Context Protocol (MCP).
It is designed as a lightweight, self-hosted alternative to OpenClaw, with a strong focus on:
- Local LLM execution (Ollama-based)
- Structured tool invocation via MCP
- Modular persona-driven development agents
- Minimal cloud dependency
- Transparent and hackable architecture
mini-claw is built on three core ideas.
-
Local First All model inference runs on local LLMs (e.g., Ollama). No external API dependency is required.
-
Explicit Tooling All capabilities are exposed through MCP servers. Agents do not act implicitly. They call tools.
-
Composable Agents Agents are LLM + prompt + tool access + orchestration logic.
mini-claw is not just automation. It is an experiment in building controllable, inspectable AI systems.
User ↓ Orchestrator (Agent Loop) ↓ Local LLM (Ollama) ↓ MCP Tool Invocation ↓ Tool Execution ↓ Structured Response
Agent Orchestrator Controls the reasoning loop and tool execution cycle.
Local LLM Backend Runs via Ollama. Responsible for reasoning and tool selection.
MCP Servers Expose tools such as:
- File manipulation
- Code execution
- QA / testing
- Development workflows
Personas Role definitions that shape behavior (developer, reviewer, tester, etc.)
- Build a reproducible local development swarm
- Understand deep MCP mechanics
- Explore agent autonomy safely
- Create a foundation for future distributed AI workflows
- Cloud dependency
- Hidden automation
- Opaque decision-making
git clone https://github.com/ndev-hoster/mini-claw cd mini-claw python -m venv env source env/bin/activate pip install -r requirements.txt
Ensure Ollama is installed and a model is available:
ollama pull llama3
Early architecture phase. Under active development.
mini-claw evolves concepts prototyped in the intro-to-mcps repository and formalizes them into a cleaner, production-oriented structure.
TBD