Local-first AI agent that edits its own output in a browser

What it does

Lemon AI is a full-stack agent framework designed to run entirely on your own hardware using local LLMs through Ollama or VLLM. It handles research, coding, data analysis, and content creation inside a Docker-based VM sandbox meant to isolate file system and OS access. The project also ships with a visual editor for tweaking AI-generated HTML pages by clicking individual elements and having the agent rewrite just those sections.

The interesting bit

Most agent frameworks generate a blob and hand it back to you; Lemon AI’s editor treats the output as a living document you can patch iteratively without trashing the rest of the page. The README is light on architectural specifics, but the sandboxing approach and the click-to-edit loop suggest the authors are trying to solve the “one wrong edit ruins everything” problem that plagues vibe-coding workflows.

Key highlights

• Runs fully offline using local models (DeepSeek, Qwen, Llama, etc.) via Ollama or VLLM, with optional cloud API fallback
• Docker VM sandbox isolates code writing and execution from the host machine
• Built-in HTML editor supports AI-driven section edits, insertions, reformatting, and direct manual tweaks
• Claims a “self-evolving” memory system that personalizes behavior across conversations
• Deployable as open-source code, container, client app, or online subscription

Caveats

• The README makes aggressive cost claims (“1/10 to 1/100 of other agent products”) without supporting data or methodology
• Several sections read like marketing copy rather than technical documentation, and details on the sandboxing mechanics and memory implementation are thin
• The license is described as “Apache 2.0 with a few additional restrictions,” but the README does not spell out what those restrictions are

Verdict

Worth a look if you want a local, browser-friendly agent for research and vibe-coding with iterative output refinement. Skip it if you need rigorous technical documentation or a proven, battle-tested sandbox before letting an LLM touch your file system.