SmallClaw is a local-first AI agent framework designed for small language models, offering a cost-effective, privacy-preserving alternative to cloud AI assistants. It empowers users to run sophisticated agents on their own hardware, providing tools like file manipulation, web search, and browser automation via a web UI. The core benefit is enabling powerful AI functionality without recurring API costs or data privacy concerns.

How It Works

The architecture features a novel single-pass chat handler. Each message triggers one LLM call to decide between conversational response or tool execution, drastically reducing latency and optimizing performance for small models. It leverages native tool-calling formats (e.g., Ollama's) for reliable structured output, avoiding free-form code generation issues. Tool results are streamed back into the conversation for iterative response refinement.

Quick Start & Requirements

Installation involves npm install and npm run build, followed by npm start or smallclaw gateway start. Prerequisites include Node.js 18+ and a configured model provider (Ollama, llama.cpp, LM Studio, OpenAI API). Minimum 8GB RAM is required (16GB recommended for coding). Docker Compose is also supported.

Highlighted Details

• Precise file read/write/edit operations.
• Multi-provider web search and full-text fetching.
• Playwright-powered browser automation (navigation, forms, screenshots).
• Extensible skills system via SKILL.md.
• Optional multi-agent orchestration.
• Optimized for small models (short history, native tool-calling).

Maintenance & Community

Developed by Xpose Market under the MIT license. Donation links are provided, indicating community support. Recent updates address bugs in browser automation and task resumption, showing active development.

Licensing & Compatibility

Released under the permissive MIT license, supporting broad commercial and closed-source integration. Its local-first design ensures data privacy and avoids cloud vendor lock-in.

Limitations & Caveats

Version 1.1 indicates ongoing development. Key missing features include persistent sessions, a background task daemon, and integrated Git operations. Optimization is focused on small models, potentially requiring tuning for larger or specialized LLMs.