ChatALL

Last reviewed Aug 26, 2026 · 707 words

ChatALL is a desktop app, not a server, and the most useful thing to understand before installing it is that it earns its keep for exactly one job: typing a prompt once and reading 5 to 10 answers side by side, from hosted services and from your own local models at the same time. Nothing else in the self-hosted AI toolbox does that comparison view well. For daily chat it is the wrong tool, and Open WebUI is better on every other axis.

It lives on your laptop, and that is the right place for it

There is no container to deploy. ChatALL is an Electron application (JavaScript, Apache-2.0, 16,482 stars) for Windows, macOS and Linux that idles around 512 MB of RAM. The "self-hosted" angle is entirely about what it talks to: hosted bots such as ChatGPT, Claude and Gemini are reached either through their APIs with your keys or through embedded web sessions you log into inside the app, and local models are reached over HTTP on your LAN. Everything you type stays on your machine except the copies sent to each bot you have enabled, which is the point to remember when you paste something sensitive with 8 bots switched on.

The interface is a grid. Each column is a bot, each row is a prompt, and you toggle bots on and off per session. There is a light thread history and a settings page per bot, and that is the whole product. Simple is the strength here.

Wiring in your own models

The interesting configuration is the local column. Ollama is the straightforward option: run it on any box on the network, set OLLAMA_HOST=0.0.0.0:11434 in its environment so it listens beyond localhost, and give ChatALL the base URL and a model name. Any OpenAI-compatible endpoint works through the generic OpenAI bot with a custom URL, which covers LM Studio, llama.cpp's server, vLLM and LocalAI. Add two or three quantisations of the same model as separate bots and you have a quick way to see whether the 4-bit build actually answers worse than the 8-bit one on your prompts.

That comparison is the real reason a self-hoster wants this. Before committing a GPU and a weekend to running a model, put it in a column next to a hosted frontier model and your current favourite, run 20 of your real prompts through, and read the answers together. It is not a benchmark, and the local LLM benchmarking guide explains why anecdotes and eval scores diverge, but it is the fastest sanity check available.

Where it breaks

The web-session bots are the fragile part. They rely on the provider's web front end staying stable, and providers change front ends and tighten anti-automation checks regularly, so a column that worked last month can sit blank until the next ChatALL release. API-key bots do not have that problem, but they cost money per token. Some services will not work without an account and a few block automated access outright.

The other limits are structural rather than bugs. There is no retrieval over your documents, no shared history between machines, no multi-user access and no mobile client. It does not manage prompts or system messages with any sophistication. If you want any of that, the AI chat interfaces category is where the server-side tools live, and Open WebUI or LibreChat with a model-comparison side panel gets you 70% of the way to what ChatALL does while also being a proper chat client.

What I'd do

Install ChatALL on the desktop, connect Ollama on the homelab plus one or two API keys, and use it the way you would use a scratch notebook: prompt bake-offs, quantisation checks, and settling arguments about which model is better at the thing you actually do. Keep Open WebUI as the daily driver and the place where history and documents live. Do not try to make ChatALL a shared team tool; it was never designed for that and the desktop-only shape is the honest signal.