Local autonomous AI agent, browses the web, extracts data, and executes tasks without any API key.
| Agent browsing the web | Knowledge Base (RAG) |
|---|---|
![]() |
![]() |
An autonomous AI agent that runs 100% on your machine. No accounts, no tokens, no API keys.
It combines a local LLM (Gemma 4 via Ollama) with a real browser (Playwright) to perform web tasks: scraping, real-time research, data extraction.
You send a task
↓
Gemma 4 decides which tool to use
↓
Playwright opens the browser and loads the page
↓
The result is sent back to Gemma 4
↓
Streaming response with cards showing each step
Requirement: Docker Desktop
git clone <url> yk-agent
cd yk-agent
cp .env.example .env
docker compose up -dOpen at http://localhost:3001
On first run, the model is downloaded automatically (~2.5 GB for
gemma4:e4b). Monitor progress withdocker compose logs -f model-init.
Go to https://example.com and tell me what's on the page.
Go to https://news.ycombinator.com and give me the titles of the top stories.
Go to https://github.com/ollama/ollama and summarize the project.
- Autonomous browsing with live tool cards showing each action (Navigate, Snapshot)
- Multiple sessions with history saved locally
- Knowledge Base (RAG): add URLs, text snippets, or PDFs; the agent uses that content with a badge showing how many sources were used
| Model | RAM | Best for |
|---|---|---|
gemma4:e2b |
~1 GB | Low-memory machines |
gemma4:e4b |
~2.5 GB | Recommended |
gemma4:12b |
~7 GB | More complex tasks |
llama3.2 |
~2 GB | Lightweight alternative |
Edit GEMMA_MODEL in .env to switch. Any tool-calling capable Ollama model works.
docker compose logs -f # live logs
docker compose ps # service status
docker compose down # stop
docker compose down -v # stop and delete all datadocker compose logs yk_agent_api # agent not responding
docker compose logs yk_playwright # browser not connecting
docker compose logs yk_model_init # model not downloadedMIT License, YellowKode + Wunka Tech

