A developer built one web page to replace five AI chat tabs

The problem: too many tabs

As dev.to reported, a developer describes an AI workflow that had turned into a part-time job. Claude Code ran in one terminal, Codex in another, a local Llama model ran through Ollama, and DeepSeek sat in a browser tab that kept getting lost. Every task began with the same question: which window was I in, and where did that conversation go?

The author first tried OpenWebUI, an existing tool meant to bring such things together. By their account, an afternoon of trying to connect an agent produced a lot of configuration text and no working agent. So they teamed up with two friends and built their own, as an entry in the Hacktoberfest Open-Source AI Challenge, Week 3.

What AgentWebUI actually is

AgentWebUI is a web page you run yourself, on your laptop or a home server. That is what self-hosted means: the software lives on your machine, not in a company's cloud. You open it and get one sidebar listing every model and agent you use.

An agent here means a coding assistant that can take a task, work on files in a folder and report back, rather than just answer questions. The home screen asks whether you want to ask something or hand a task to an agent. The author says agents are the main focus and chat is the supporting feature, the reverse of most chat tools.

In the demo screenshot, the sidebar holds a DeepSeek chat, a Claude Code workspace, Codex, OpenAI and Gemini in one list. The model picker shows llama3.2:3b running locally.

How it connects things

The README describes two kinds of connection. For models, it accepts any service that follows the OpenAI-compatible format, a widely copied way of talking to AI models. The list includes OpenAI, OpenRouter, NVIDIA, Groq, Ollama, LM Studio and vLLM.

For agents, it uses the Agent Client Protocol, a shared language that lets command-line agents talk to other programs. Claude Code, Codex, Gemini CLI and OpenCode are among the agents named. You pick the agent, then its model and effort level, then the folder it should work in. The README says it shows the agent's thinking, plan, tool calls and file changes as they happen, and asks before it acts.

Claude Code and Codex can run on the subscriptions you already pay for, so there is no separate key to set up. Everything is controlled by an env file, a plain text file of settings, so moving from laptop to server is mostly a copy.

Small on purpose

The most concrete claims are about size. The README says the server is about 3,000 lines of Python, depends on one outside library (aiohttp) and uses about 40 MB of memory. The web page itself uses Preact, a small interface library, with no build step. Nothing loads from the internet unless you choose a font other than Geist.

Setup is three commands and a settings file. You download the code, copy the example env file and fill in your models and agents, then start it with Docker Compose. Docker is a tool that packages software so it runs the same way on different machines. The README says there is no setup screen and no accounts. The package and command are still named agentwebui-lite.

The local-first argument

The author makes a case for open models in this setup. A local Llama answers even when the Wi-Fi is down, and quick questions stay on the laptop. Agents work on your own projects on your own server. When a new open-weight model appears (one whose trained settings are published for anyone to run), you change the env file and restart.

The author also suggests a cost split: small jobs go to free local models, and the paid subscriptions handle only heavy work. That is a personal workflow claim, not a tested result. The author's own comparison table is described as a study of their own laziness, so treat it as a joke.

How much weight to give it

This is a side project from a challenge, written with a lot of humor. We have not run it, and the claims come from the author's post and project page. The author also opens with a story about the chainsaw's medical origins to argue that laziness drives invention.

Still, the problem is real. Anyone using several AI services knows the scattered history and repeated logins. Whether this particular tool lasts matters less than the pattern: a thin, self-hosted layer that lets you swap models without rebuilding your habits.