AgentLLM is a proof-of-concept for browser-native autonomous agents, targeting researchers and developers interested in on-device LLM capabilities. It demonstrates that embedded LLMs can handle complex goal-oriented tasks with acceptable performance, offering a privacy-preserving and cost-effective alternative to server-based agents.

How It Works

AgentLLM leverages WebLLM and WebGPU to run LLM inference directly in the browser, utilizing the GPU for significant performance gains over CPU-based methods. It modifies the AgentGPT project, replacing ChatGPT with WizardLM and altering the prompt mechanism. This approach allows agents to perform arbitrary goals by generating and executing tasks in a loop, without external tools, simplifying complexity and providing a user-friendly GUI for rapid prototyping.

Quick Start & Requirements

• Demo: Install Chrome or Edge (v113+). Launch with --enable-dawn-features=disable_robustness for performance. Navigate to the AgentLLM web interface.
• Docker: Run ./setup.sh --docker or ./setup.sh --docker-compose.
• Local Development: Requires Node.js 18+. Clone repo, npm install, create .env (with NEXTAUTH_SECRET, NEXTAUTH_URL, OPENAI_API_KEY), run ./prisma/useSqlite.sh (optional), npx prisma db push, and npm run dev.
• GitHub Codespaces: Click "Code" > "Codespaces" on the repo, create a Codespace, run ./setup.sh --local, add API key, and open in browser.

Highlighted Details

• Browser-native LLM inference via WebGPU.
• Based on AgentGPT architecture, using WizardLM.
• Focus on goal-oriented task generation and execution loop.
• Proof-of-concept for research, not production-ready.

Maintenance & Community

• Project is a proof-of-concept by idosal.
• Links to Medium and Twitter for updates.

Licensing & Compatibility

• License not explicitly stated in the README.
• Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

This project is a proof-of-concept utilizing experimental technologies and is not production-ready. Performance may vary significantly based on device capabilities, with lower-tier devices potentially unable to run the demo. An OpenAI API key is required for setup, despite the focus on browser-native LLMs.