Self-host Ollama GUI

Ollama GUI is a browser chat front end for models served by an Ollama instance on your own hardware, suited to people who want private LLM conversations without a vendor account. The project ships no server component of its own: it is a single-page app whose conversation records sit in browser storage (IndexedDB), and its only network peer is the Ollama endpoint you configure. Answers arrive as a stream you can halt, redo or amend, with amended turns preserved as switchable versions and any reply branchable into a fresh chat. Thinking-capable models such as qwen3 and deepseek-r1 surface their reasoning in a foldable panel, while gpt-oss exposes an effort setting; vision models accept image uploads, and text or code files can be attached to a message. A bundled snapshot of the model catalogue lets you look up models, pull them with progress feedback, remove them, read their capabilities and context length, or drop them out of RAM. System prompts and sampling options like temperature, top_p, num_ctx and seed can be configured app-wide, per model or per conversation. Older threads are ordered by day and support pinning, archiving, renaming, automatic titles and keyword search; rendered Markdown carries code highlighting plus KaTeX equations, and everything exports to JSON or a single thread to Markdown. Self-hosting is one command: the ghcr.io image runs nginx serving the site and forwarding /api to Ollama, the compose file adds the model server itself at http://localhost:8080, or you can build the static bundle with Bun and reverse-proxy it. The hosted build at ollama-gui.vercel.app reaches a local Ollama once that origin is allowed. History is scoped per origin, so another port, browser or private window shows an empty list, and over plain HTTP Gravatar pictures and offline caching are unavailable.