Sharing a project I maintain. WebBrain puts an agent in the browser side panel: it can summarize/answer questions about the current page, extract tables and lists, fill forms, and navigate, using whichever model you configure (llama.cpp/Ollama/LM Studio locally, or any OpenAI-compatible API).
Some things that might interest people here:
- Plain vanilla JS, same codebase shipped for Chrome and Firefox
- Works with small local models: context trimming, a compact prompt mode, and a narrow tool set
- New MCP server, so Claude Code / Codex / OpenCode can use your real signed-in browser via the extension
Contributions welcome, especially site adapters and bug reports with a URL + the task that failed.
How is giving something, that is nondeterministic, access to your web browser a good idea? How do you guarantee that it doesn’t enter sensitive information without the user’s explicit consent? How do you guarantee privacy and security?
I’m skeptical, but I am genuinely asking and willing to listen.