zeithub.otto: an AI IDE on Electron and Next.js for any models

⚡ The gist in 5 seconds - zeithub.otto is a desktop AI IDE on Electron and Next.js that connects any models: local ones via Ollama, cloud providers (OpenRouter, Groq, NVIDIA, Cohere, Z.ai, Cloudflare), as well as Claude Code and Codex. - A Windows installer is available on GitHub; it can also be built from source (Node.js 20+ required). - Limitation: the installer is not yet code-signed, so Windows SmartScreen may warn — you need to choose "More info → Run anyway". ### 🔍 What was found The zeithub-team/otto repository has published the "zeithub.otto" project — a desktop AI coding studio designed to work with models of various levels: from a 4B model on a regular CPU to 35B on a gaming GPU, including Claude and Codex in the cloud. The architecture is split into an Electron shell (apps/desktop) and a Next.js interface (web), with a backend on Node.js and SQLite. The server listens only on 127.0.0.1, and the Electron renderer has no direct access to Node. The key feature is a guided mode for weak models: the model responds with strict JSON validated against a schema, one step at a time, and receives only the necessary tools. Thanks to this, a model like 9B can write real files rather than pasting code into chat. When a provider is unavailable or out of quota, Otto automatically switches to the next provider, and finally to a local model, showing a "switched from A to B" note instead of an error. After generating a site, the IDE opens it in a headless browser (at desktop and mobile widths) and sends the model a list of problems: overlapping blocks, unstyled forms, default blue links, horizontal scrolling. Also claimed: step-by-step building of multi-page sites, file corruption protection (e.g., from unclosed brackets), app control via the model, a built-in terminal, Tabby, device previews, attachments of various types, a skills library and an interface in six languages. ### 💡 Why it matters zeithub.otto solves a problem faced by all modern AI editors: they are built for strong models, and when a small local model is connected they start hallucinating, forget to save files or get stuck in loops. Otto was built the other way around — so that local models work in a controlled mode, while large ones run without restrictions. Practical benefit: you can bring your own model via Ollama (qwen3.5:4b, gemma4:12b, qwen3.6:35b and others are recommended) and see how the IDE handles tasks without cloud costs. In a "coffee shop" test (one landing page, four follow-up requests, browser check), gemma4:12b scored 100/100 in under 2.5 minutes. ### 🧩 Context The project is publicly available on GitHub under the MIT license and belongs to the zeithub.team team. The repository describes implementation details: the chat loop, fallback between providers and "nudging" models that describe work instead of doing it, are in src/server/ollama.ts; the guided mode with JSON schema is in src/server/guided.ts; site generation and browser-based checking are in sitebuilder.ts and visu

🔗 Read on github.com